United States
              Environmental Protection
              Agency
              Environmental Sciences Research  EPA-600/4-79-035
              Laboratory          May 1979
              Research Triangle Park NC 27711
              Research and Development
vvEPA
Systematic         !;
Sensitivity
Analysis of Air    -
Quality Simulation
Models

-------
                RESEARCH REPORTING SERIES

Research reports of the Office of Research and Development, U S Environmental
Protection Agency, have been grouped into nine series These nine broad cate-
gories were established to facilitate further development and application of en-
vironmental technology  Elimination of traditional grouping  was  consciously
planned to foster technology transfer and a maximum interface in related fields.
The nine series are'

      1   Environmental  Health Effects Research
      2   Environmental  Protection Technology
      3   Ecological Research
      4   Environmental  Monitoring
      5   Socioeconomic Environmental Studies
      6   Scientific and Technical Assessment Reports (STAR)
      7   Interagency Energy-Environment Research and  Development
      8.  "Special" Reports
      9.  Miscellaneous Reports

This  report has been assigned to the ENVIRONMENTAL MONITORING series.
This  series describes research conducted to develop new  or improved methods
and  instrumentation for the identification and quantification of environmental
pollutants at the lowest conceivably significant concentrations. It also includes
studies to determine the ambient concentrations of pollutants m the environment
and/or the variance of pollutants as a function of time or meteorological factors.
This document is available to the public through the National Technical Informa-
tion Service, Springfield, Virginia  22161.

-------
                                             EPA-6QO/4-79-Q35
                                             Way 1979
       SYSTEMATIC SENSITIVITY ANALYSIS

       OF AIR QUALITY SIMULATION MODELS
       Robtert J, Geltnas and J. Peter Vajk-
          Science Applications, Inc,
        1811 Santa Rita Road, Suite 104
         Pleasanton, California  94566
           Contract No. 68-02-2942
               Project Officer
             Robert E, Eskridge
     Meteorology and Assessment Division
 Environmental Sciences Research Laboratory
Research Triangle Park, North Carolina  27711
 ENVIRONMENTAL SCIENCES RESEARCH LABORATORY
     OFFICE OF RESEARCH AND DEVELOPMENT
    U.S. ENVIRONMENTAL PROTECTION AGENCY
RESEARCH TRIANGLE PARK, NORTH CAROLINA 27711

-------
                              DISCLAIMER
     This report has been reviewed by the Environmental Science Research
Laboratory, U.S. Environmental Protection Agency, and approved for
publication.  Approval  does not signify that the contents necessarily
reflect the views and policies of the U.S. Environmental Protection
Agency, nor does mention of trade names or commercial products constitute
endorsement or recommendation for use.
                                   11

-------
                            ABSTRACT
     This report reviews and assesses systematic sensitivity and
uncertainty analysis methods for applications to air quality
simulation models.   The discussion of the candidate methods
presents their basic variables, mathematical  foundations, user
motivations and preferences, computer implementation properties,
and costs (including both human and computer  resources).   Both
deterministic and sampling methods have been  evaluated with
illustrative examples of "What you pay"and "What you get" with
present-generation  systematic sensitivity analysis methods and
air quality models.   Deterministic methods include the time- and
user-honored method  of arbitrary parameter adjustments by trial
and error, variational methods, and a newly formulated Green's
function method (for kinetics systems only).   Sampling methods
include Monte Carlo  sampling of the outputs of air quality
models to compute variances and a Fourier analysis method of
sampling model outputs to compute expectation values of sensitiv-
ity coefficients.

     Computational  economics, inclusive of both programming
effort and computer  execution costs, are the  dominant governing
factors for the effective application of systematic sensitivity
and uncertainty analysis methods to arbitrary air quality
problem scenarios;  several reasonable options and several un-
reasonable options  emerge from the present evaluations.  Recommen-
dations outline how  the EPA should, today, most effectively
apply available multi-parameter systematic methods of sensitivity
analysis to arbitrary 3-D transient (PDE) air quality simulation
models.  The report  concludes with a discussion of key break-
throughs which would lead to the next major advances in system-
atic sensitivity analysis.  In certain instances, these would
simultaneously provide major advances in air  quality simulation
modeli ng , per se.

     This report was submitted in fulfillment of Contract
No. 68-02-2942 by Science Applications, Inc., under the sponsor-
ship of the U.S. Environmental Protection Agency.  This report
covers a period from May 12, 1978, to November 12, 1978,  and
work was completed  as of December 4, 1978.

-------
                                CONTENTS

Abstract	iii
Figures	    vi
Tables	vii
Acknowledgments  	  viii
   1.  Introduction  	     1
   2.  The Problem	     6
            Air Quality Models and Their Sources of Uncertainty  .  .     6
            Motivations for Sensitivity Analysis 	    12
            Analytical Basis of Sensitivity Analysis Methods ....    14
   3.  Systematic Sensitivity Analysis Methods 	    30
            Deterministic Methods  	    30
            Sampling Methods 	    48
   4.  Evaluation of Candidate Sensitivity Methods for Air
       Quality Model Applications  	    61
            Basic Cases	    63
            Chemical Model Evaluation  	    66
            1-D Time-Dependent Model Evaluation  	    76
            3-D Air Quality Model Evaluation 	    82
   5.  Today's Approach  	    85
            Conclusions	    85
            Today's Practical Application of Systematic
              Sensitivity Analysis 	    87
            Needed Breakthroughs in Systematic Sensitivity
              Analysis Methods	    92
References	    95

-------
                                 FIGURES

Number                                                              Page
  1   Effects of parameter variations on the solutions 	   19
  2   Schematic representation of solution surfaces  	   22
  3   Interpretation of probability distributions of
        uncertain parameters 	   24
  4   Construction of first-order sensitivity measures and
        coefficients 	   26
  5   Schematic illustration of the Monte Carlo method 	   49
  6a  Schematic illustration of the Fourier Amplitude
        Sensitivity Test  (FAST)	   52
  6b  Functional output of FAST	   54
  7   Arbitrary nature of functional forms used in FAST   	   56
                                    vi

-------
                                 TABLES

Number                                                              Page
  1   Characteristics of candidate systematic sensitivity
        and uncertainty analysis methods 	    60
  2   Base cases for illustration of execution costs of
        candidate sensitivity analysis methods 	    83
  3   Illustrative computer execution costs for candidate
        systematic sensitivity analysis methods  	    84
                                  vii

-------
                 ACKNOWLEDGEMENTS
     It is a pleasure to thank Dr.  Richard  Stolarski  and
Prof.  John Seinfeld for in-depth discussions  of their
sensitivity work prior to publication.   We  are  also
extremely grateful  to our colleague,  Dr.  Said Doss, for
his exacting-review and suggestions regarding some of
the complex issues  in numerical  analysis  which  were
encountered in the  course of this  study.
                           VI 1 1

-------
                       IN
                          ECTION 1
RODUCTION
    Uncertain parameters  and sparse  field data  are

intrinsic elements  of air quality modeling.   This state

of affairs will  persist indefinitely,  despite the nation's

ever more urgent need for more detailed understanding  of

cause and effect mechanisms in air quality.   Thus a  key

question is, "To what extent do uncertain input factors  in

air quality models  lead to impaired  accuracy and precision

of the resultant outputs  of these models?"  This basic

question and many of its  variants are  examined  in this final

report on "Systematic Sensitivity Analysis for  Air Quality

Simulation Models ".


    The principal objectives of this study were:


          (1)  to identify those systematic sensitivity
               analysis methods which  are applicable to
               air  quality simulation  models;

          (2)  to evaluate critically  the potential  of
               candidate  methods of  systematic  sensitivity
               analysis for effective  application to key
               air  quality issues;

          (3)  to identify those breakthroughs  most critically
               needed for each candidate method; and

          (4)  to determine the most effective  immediate
               means of conducting sensitivity  analyses  on
               existing air quality  simulation  models,
               assuming no further developments in sensitivity
               analysis methods.

-------
    It is frequently unrecognized that sensitivity analysis
is a basic ingredient in  any application of the classic
scientific method in which  alternative hypotheses--including
models and input data--are  tested and retested against
selected physical realizations.   Historically, this process
has most frequently been  conducted on a trial  and error
basis.  In the development  of air quality models, sensitivity
testing is a particularly laborious process if all potentially
significant input parameters are to be varied, individually
or simultaneously, for every potentially significant spatial
1ocati on.

    Somewhat more systematic approaches to sensitivity analysis,
including some semi-automated methods, have appeared recently
in the study of purely chemical  kinetics systems.  Enormous
benefits would undoubtedly  be realized in air qual-ity analysis
and policy making if semi-automatic sensitivity methods could
be extended to economic applications in arbitrary space- and
time-dependent multivariate models.  Such extensions of
systematic sensitivity analysis  thus represent both a challenge
and a promise for the entire field of air quality work.  The
principal focus of this study is the evaluation of this duality
--the challenge versus the  pnomise--in the application of
candidate sensitivity analysis methods to arbitrary air quality
models.

    Perhaps the key consideration  for air quality applications
is the  computational economics of  candidate sensitivity
methods.  On the  one hand,  we shall see that  several dedicated
methods would require that a specific sensitivity program be
written for each  specific application to an air quality model.
On the other hand, we shall see that several  non-dedicated
methods can be applied, without additional programming effort
for either the air quality model or the general sensitivity
                         2

-------
 method,   to  any  air  quality  model.   All  sensitivity
 analysis  methods  considered--in  their current state
 of  development--wi11  be  seen to  have both  strengths
 and weaknesses  in  regard to  potential applications
 to  1-D,  2-D,  and  3-D time-dependent  air  quality  models.
 The extent  to which  the  weaknesses will  be disabling
 and/ or  the  extent  to which  the  strengths  can be ex-
 ploited  will  necessarily depend  upon the needs and
 expectations  of  potential  users  of these systematic
 methods.   The motivations  of air quality policy  makers
 and of air quality researchers,  while closely coupled,
 are sufficiently distinct that they  may  be served best
 by  different methods of  viewing  uncertainties and sen-
 si ti vi ti es.

     In the  present  study,  it has been necessary  to  em-
 phasize  issues  relating  to the sensitivities of  air  quality
 models With  respect to those physical, chemical, or  meteorological
 factors  which are  most readily quantifiable.  Thus we  have
 considered  variations in model outputs — such as  the  concen-
 trations  of  Oj.  NO,  N02»  and  other air-borne species — which
 result from  uncertain and variable  input factors — such  as
 photochemical rates, emission sources, pollutant sinks,  and
 winds.  We  have  not  considered sensitivities to  computational
 factors  since no practical standards exist for comparative
 analysis.   It would  be impossible,  for example,  to  determine
 the dispersion  in  oxidant concentrations attributable  to
 numerical  diffusion  and/or to operator-splitting in  the
 Systems  Applications, Inc.,  regional model^ ' vis-a-vis
 the LIRAQ model'  '  vis a vis the MADCAP  model^ for a  specific
"typical  day"  in,  say, St.  Louis, even if the same data  bases,
initial conditions,  and boundary  conditions were  incorporated

-------
in each model.  The multitude of other uncontrolled differ-
ences among air quality models precludes meaningful con-
sideration of model-to-model sensitivities at this time.
    We have thus surveyed and evaluated those sensitivity
analysis methods which can, for any given air quality model,
determine systematically and quantitatively the dispersion
in the output variables--such as space- and time-dependent
concentrations of ozone, PAN's, NO ,  acids, aldehydes,
                                  J\
peroxides, and SO -- which result from uncertain photo-
                 X
chemical rate data, winds, diffusivitives, initial and boundary
values, and emission sources and sinks, and from deviations
of actual conditions from "typical day" conditions.

    Four types of systematic sensitivity methods are discussed
in this report:

    (1)  "time-honored methods," which examine variations
         of selected parameters, either individually or
         simultaneously, by trial and error;
    (2)  variational methods (including direct methods);
    (3)  Fourier  amplitude sensitivity test  (FAST) methods;
    (4)  Monte Carlo methods.
    The report is presented in four sections.  In  Section
2, The  Problem, we  discuss the motivations,  logical  bases,
and key issues involved  in the application of  systematic
sensitivity analyses to  air quality simulation models.   In
Section 3, Systematic Sensitivity Analysis Methods,  we de-
scribe  the candidate methods in  terms of  their basic variables,
mathematical  formulation,  computer implementation, and execu-
tion  properties.

-------
      In Section 4, Evaluation of Candidate Sensitivity
 Methods for Air Quality Model Applications, we discuss
"what  you pay" , "what you get", areas of promise, and
 problems.  Section 5, Today's Approach, indicates the most
 effective  means,  in  our opinion, of applying  systematic
 sensitivity methods  to air  quality models  in  the  face
 of  the  current adverse constraints of  computer economics.
 Finally, we indicate  some  of  the needed breakthroughs
 which would lead  to  dramatic  advances  in  the  practical
 utility of candidate  sensitivity methods.

-------
                         SECTION  2
                        THE PROBLEM

AIR QUALITY MODELS AND THEIR SOURCES  OF  UNCERTAINTY
    Air quality simulation models attempt to describe the
physical, chemical, and photochemical  processes occurring
in specific portions of the lower atmosphere.   Variables
of principal interest include the concentrations of  sus-
pended particulate matter and of  such gas phase chemical
species as ozone, nitrogen oxides,   and   carbon monoxide,
for which regulatory standards have been set by various
governmental jurisdictions.  Although no such regulatory
standards have been set for species such as aldehydes,
peroxides, nitrates, and acids, these and other species
are also the subject of extensive current interest.   The most
thoroughly developed and calibrated air quality models
available today describe, for the most part, the photo*
chemical kinetics and the transport processes for gas
phase  species  in  certain limited atmospheric regions.

    Several dissimilar  sources of uncertainty in our know-
ledge  of atmospheric processes affect the predictions of
air quality simulations.  Such uncertainty  factors include:

 (i)   Natural  Variabilities
           •  winds                 t  temperature
           0   solar  flux           •  atmospheric turbulence
           e   humidity              •  atmospheric stability
           •   natural  sources  and   •  particulate matter
              sinks  of  ai r-borne
              chemical  species

-------
      It is  important to note here that significant
natural  variabilities occur in the real atmosphere
over scales  which are often disparate with respect
to the scales used in practical  computations.  For
tropospheric models,  typical  computational scales
are tens of  meters to tens or hundreds of kilometers.
Mixing of chemical species, on the other hand,  occurs
on microscopic scales comparable to molecular  mean
free path lengths.  Small  scale  eddies and chemical
mixing are thus, in most instances, treated by  sub-
grid scale characterizations  rather than by detailed
calculations of reactive turbulence.   Winds at  specific
grid points  in air quality models are generally inter-
polated from discrete field data from monitoring stations
separated far wider in space  and in time than  the horizontal
or vertical  grids or integration time-steps of  the models.

      Variable attenuation and multiple scattering
effects associated with particulate matter, gaseous
components,  cloud cover, and  surface albedo all contribute
to significant local  variations  in solar flux.  Coastal
areas, in particular, show extreme variations  within a
single geographic area and within a single day.

      Naturally occurring chemical sources include forest
fires, ocean spray, dust,  local  geogenic emissions from
oilfields and coalfields,  eutrophication of inland waters
and oceanic  processes releasing  hydrocarbons,  N^O, CO,
NH3, and H2S, as well as a nearly universal background
release of methane.  To these are added widespread
sources attributable to agriculture and deforestation
around the globe.  Natural chemical sinks include hetero-

-------
    geneous  reactions with  suspended  participate matter,
    rainout,  absorption  at  the  air/ocean  interface, and
    other  surface  deposition  processes.

    (i i )   Photochemical  Factors
            0  upper  and lower  atmospheric  compositions
               (are  determinants  of solar flux  intensity
               and spectral  distribution)
            e  diurnal  variations  in  solar  flux
            o  absorption  and photodissociation cross-
               section  data
            o  hydrocarbon  emission inventories
            o  photochemical  characterizations  of  aggregate
               hydrocarbon  emissions*

          Photodissociation  rates  are described by wave-
    length integrations  over  products of  local  solar  flux
    and photodissociation  cross-sections.   These rates  thus
    depend strongly  on  the  spectrum of the  local solar  flux
    which  is  determined  by  atmospheric composition and  time
          (6)
    of dayv   .   Because  the  concentrations  of minor reactive
    species--such  as  OH, 0,  HUO^,  NO-,, and  N20g--have seldom
    been  measured  simultaneously  with solar flux and  the  con-
    centrations of other important atmospheric  species, much
    of the validation of photochemical mechanisms  remains  to
    be done.
          Atmospheric hydrocarbon inventories,  of  course,
    consist of huge  numbers  of individual  organic  compounds
    which must be  aggregated according to  their reactive
    properties in  a  photochemical mechanism.   Additionally,
*See,  for example,  the articles of Hecht,  Seinfeld,  and
DodgeU), Hogo and  Whitten (5), and Gelinas  and Skewes-Cox(6)
for detailed characterizations and categorizations  of reactive
hydrocarbons.

-------
detailed emission inventories and local hydrocarbon
distributions are difficult and costly to obtain for
regions as extended as those spanned by air quality
simulation models.

(i i i)   Field Data

     Measurements of such meteorological variables
as  wind, temperature, stability, and transported
chemical species are relatively sparse in both time
and space.  Because these quantities are associated
with  exceedingly complex fluid dynamics processes
they  have usually not been calculated from fundamental
equations but have instead been imposed upon air quality
simulation models using interpolations of the measured
data  onto the spatial grids and time points of the
model.   This imposition introduces uncertainties which
are associated with the specific interpolation processes
as  well as uncertainties which are associated with non-
representative measuring stations and with the measure-
ment  methods per se.  An additional form of uncertainty
arises  because meteorological conditions are necessarily
categorized to represent conditions which  occur on
selected "typical days".  Actual meteorological conditions
on  any  specific day will naturally differ from these
"typical day" conditions, thus constituting yet another
possible source of dispersion in the outputs of air
quality simulation models.

-------
(i v)   Computational  Factors

     Air quality simulation  models are formulated from
non-linearly coupled partial  differential  equations in
space and time describing physical transport and chemical
kinetics processes.   Very simple models are sometimes
derived from systems of non-linearly coupled ordinary
differential equations in time only.  In either case,
numerical integration methods are the most effective
means for obtaining solutions of the basic modeling
equations.  Such numerical  solutions are inherently
approximate, so that simulation model results inevitably
include dispersion due to these computational (numerical)
approximations.

      Computational dispersion usually  arises from
the well-known dilemmas of numerical modeling, most
notably numerical diffusion  (in Eulerian formulations)
and aliasing of individual  species under rezoning  (in
Lagrangian  formulations).  Such computational features
as choice of differencing methods, representations of
sub-grid scale physics, simplifications in chemical
kinetics mechanisms, and imposed  profiles for pollutant
sources result in additional computational dispersions
in the  model outputs.
                       10

-------
     Model  descriptions of the processes mentioned above
are incorporated in systems of partial differential equations
(PDE's) which are generally solved by numerical integration
on large digital computers.  In the present study, we con-
sidered air quality simulation models based upon PDE's of the
form
                            =  R.(c1,c2,...,cli] + S.(x,t),
     3t                                i = 1,...N,  (2.1)

where c- = c-(x_,t) represents the mean concentration of the
i-th chemical species; _y_ is the mean fluid velocity; Kn is
a diffusion tensor; R. is the net sum of all chemical reaction
rates for production or destruction of the i-th species;
and S-(x_,t) represents the net sum of all sources and sinks
for the i-th species.   This equation is a reduced version of
more fundamental (microscopically exact) fluid dynamics
equations, and thus incorporates well-known approximations
and assumptions which are peripheral to the present discussion

     The solutions c-(x,{p},t) of Eq. (2.1) are functions of
                    I
space and time (_x, t), but they are also functions of the
parameters { p} = {y_, KD, ka, Jg, £  , c. ^t. g}. The parameters
k^ (T) represent temperature-dependent rate constants deter-
mining the thermal chemical reaction rates; J0 represents
                                             P
coefficients for the photochemical  reactions included in the
terms R-; and A  represents parameters in the  expressions for
source and sink rates included in the terms S-.   Since the
solutions for the species concentrations are functionally
dependent on the initial and boundary values,  c.  and c.
                                               '»l    '»o ,
respectively, it is often useful to include these values
in the parameter array {p} as well.
                           11

-------
     The values for this host of parameters  needed to
describe important atmospheric processes  are obtained from
laboratory experiments, field measurements,  and theoretical
or empirical prescriptions.   The source of the entire problem
under discussion in this study is the inevitable fact that
all of these parameter evaluations are uncertain to a degree.
In the face of such uncertainties, it is  natural to ask,
"For which of the parameters do the uncertainties in the
parameters produce the greatest dispersion in the model out-
puts (especially for the most important atmospheric observables)?"

MOTIVATIONS FOR SENSITIVITY  ANALYSIS

Air Q u a 1 i ty Policy

     Air pollution control agencies and other governmental
bodies or officers are  charged with the responsibility of
regulating pollutant emissions in order to meet or to main-
tain standards of air quality established for certain atmos-
pheric regions.  To the extent that models are used responsibly
in decision making, it is both fair and necessary to ask,
"What are appropriate error bars for the outputs of the
model?"  This question is appropriate whether the models used
are analytically based, detailed simulation models incorporating
as much of  the current scientific knowledge as is practical
or the models are qualitative mental models which incorporate
intuition.  In asking this question, we would like to obtain
a  statement of the un certainty--or, equivalently, the credibi-
lity—of the model results which are  used in  regulatory decisions.
The principal emphasis is then given  to determining the expected
uncertainties in important atmospheric variables, given the
manifold of uncertainties present in  the input parameters of
                           12

-------
the model.   (Analytic definitions  will  appear in later
sections of this report.)  As an example, one
might want  to know the expected uncertainty in ozone
concentrations predicted by a given simulation model at
all significant geographical  locations  and times resulting
from the combined effects of rate  data  uncertainties,
meteorological uncertainties, emission  uncertainties, etc.,
in order to assess whether proposed regulations are reason-
able and proper.

Air Quality Research

     In contrast to regulatory policy applications of air
quality simulation models in which (expected) uncertainty
analysis — or credibility analysis — is emphasized, research
applications tend to emphasize sensitivity--cause-and-effeet-
analysis.   The two concepts are, to a significant degree,
complementary.  In research applications, the scientist seeks
to determine the magnitude of the  effects which each and
every parameter input to the model produces in the outputs
of that model.  This is a more explicitly detailed approach
than that  of expected uncertainty  analysis because, in any specific
simulation  calculation, the model  must  use a single, specific
numerical  value for each parameter of the model, not a dis-
tribution  of possible values.  Best case and worst case air
quality estimates are very conveniently made in terms of indi-
vidual, unaveraged parameter variations.

     In studying the effects of varying an individual para-
meter (or  a limited set of parameters)  on the outputs of the
model, all  other model parameters  are generally held fixed,
                           13

-------
either at nominal  "best" values or at "test values"  selected
by the researcher.   (It is also possible  to consider simultane-
ous variations in  all  other parameters  in order to obtain
best or worst case  estimates.)   Sensitivity analysis is thus
used by researchers as a means  for identification of cause-
and-effect mechanisms  and for ranking various mechanisms in
order of importance in the overall system of processes deter-
mining the state of a  specific  atmospheric region.  The goals
of this approach are:   (i)  to  increase basic scientific under-
standing of atmospheric processes;  (ii)  to focus scientific
research on the most critical issues, ultimately reducing un-
certainties in existing models; and (iii) to assist regulatory
applications, both  directly and indirectly, by progress toward
the first two goals

ANALYTICAL BASIS OF SENSITIVITY ANALYSIS  METHODS

     The methods and language of statistics have been extensively
adopted in the development of systematic  sensitivity methods,
undoubtedly because interest centers on deviations in model
solutions which result from deviations  in governing input
parameters.   In terms  of air quality modeling equations which
are represented by Eq.  (2.1), a major question is, "What
variations occur in the solutions, c-(x,  (p},t), relative to
                                    1
c.j (_x,{ p°), t), when model input parameters, {p}, are subjected
to variations relative to their nominal values {p°}?"  The
input and internal  parameters for Eq. (2.1) are written most
generally in  terms of their nominal values and deviations as:
                            14

-------
               r     = r °  + >
               ci,B  = ci,B + «ci)B
               v = v° + 6v
               k   = k   + 6k
                a    a      a
                                                       (2.2)
At this  stage  nominal  values,  {p°>,  are  not necessarily
mean values,  nor are deviations, {6p},  necessarily standard
or random deviations.   Considerable  care must thus be exer-
cised to distinguish nominal  values  from mean,values or
central  values, and also to distinguish  arbitrary deviations
from standard  deviations, variances, and random variations.

    The  logical question of,  "How are nominal values to
be assigned in air quality models?",  has several answers.
In principle,  measurable parameters  such as reaction rate
data, initial  and boundary values, winds, diffusivities ,  etc
can be represented ultimately  by mean values and associated
standard deviations; it is frequently true, in practice,
that available data bases do  not allow  reliable statistical
treatment.   For example, in photochemical kinetics some
limited  number of rate coefficients  have been repeatedly
measured by independent experiments  which have been well-
prepared (in  the statistical  mechanical  sense) and thus
                          15

-------
have reliable statistical properties.  But this tends to be
the exception rather than the rule for many reactions
which are objects of current research.  Indeed, we can not
now profess to know all of the potentially important
atmospheric reaction processes, much less their mean rates
and statistical distributions.  In such cases, nominal
values of known and hypothetical reaction rates are assigned
on the basis of each individual researcher's  discretion.
Parametric deviations are also arbitrarily assigned for the
purpose of testing the sensitivity of  photochemical mechan-
isms to specific  reactions.   The history of atmospheric
photochemistry developments emphatically illustrates that
meaningful probabi1i ty distributions of poorly known reac-
tion  rates  are  simply  not  available  at times  of  greatest
need.  (Poorly known chemical rates have always been subject
to very large, frequently  unpredictable changes which emerge
when markedly new or more carefully prepared experiments are
performed.)  This discussion of meaningful probability  distri-
butions for physical parameters is particularly  pertinent  to
our  consideration  of those sensitivity and uncertainty  analysis
methods which  use  probability  distributions.

Measures of Sensitivity and Uncertainty

     Several measures are used  in sensitivity  and  uncer-
tanity analysis which should be distinguished at  the outset.

    The most elementary  measure is a  deviation  (variation)
in state variables such as species concentrations at  specific
coordinates (_.x_K,t' ) ,
    fic^xKpht1) = c.(x',{p0 + 6p},t') -  c.(x;{p°}. ,t'),    (2.3)

in the neighborhood of nominal parameter values, {p°}.
(Recall that the parameters can be either random  or non-
random vari ables. )

                           16

-------
By Taylor's theorem
                                      M
    c(x',{p° + 6p},t) = c.(x',{p°}t) + Z
     i                                 k= 1
            0((max6pk)2), -j = i,N; k = ijM,             (2.4)
where the sensitivity coefficients are defined by

    Zi'k = 8ci/9pk     >   1 = l.N; k = 1,M   .
Sensitivity methods which neglect the terms 0((max dp^)  )
are referred to as first-order or linear methods.  In  space
and time dependent models, which are described by partial
differential equations (in contrast to ordinary differential
equations for purely kinetics models or for steady state
purely fluid dynamics models), species concentration  varia-
tions associated with the k-  parameter variations  are
given by the functional derivatives,

                 p 9c, (x.1 ,t' )            .,
    6^(21',t1) = I  3P /x  tt \  <5pkU,t') djx             (2.6)

and

    6c,(x',t') =|   ^        6pk(x',t) dt.              (2.7)
        ~       J  3pk(x',t)     k


The sensitivity variable  in Eq.  (2.6)  represents  variations
in c- at (_x'»t') with respect  to  variations  in  p^  at locations
    1                                           K
x and time t1.  Likewise, the sensitivity variable  in  Eq.  (2.7)
represents variations in  c^ at (x',t') with  respect to  variations
in p. at times t and location x/.   These functional  derivatives
and the sensitivity coefficients,  Z.  ., form the  basis  of
                                    I 5 K
                           17

-------
variational  methods  which will  be described in Section 3.
These variables provide convenient measures of sensitivities
and uncertainties when probability distributions are poorly
defined and for "best case" and "worst case" air quality
considerations, which were discussed above in connection
with AIR QUALITY RESEARCH MOTIVATIONS.

    The effects of parameter variations on the solution
c-(_x,t) are shown schematically in Figure 1.  Two cases
must be distinguished here.  In Figure l(a), the solution
for the "best value" of all parameters is shown by the solid
line.  If some one of the parameters, say p., were changed
from its "best value" p°. to p*? + <$p-, the solution curve
                       J     J    J
would be different,  although it would start from the same
initial value, c- T, for c..  Likewise, if the parameter
                1,1         i
were changed to pQ -<$p., the solution curve would be different
                 J    J
again.  These altered solutions are shown by the dotted lines
in Figure l(a).  If, on the other hand, we wanted to investi-
gate the effects of changing the initial value c. T itself,
                                                i j i
the results might look like the dotted line curves shown in
Figure l(b).

    It is important to realize that it is not enough  to
consider the variations  in  the solutions ci at a single time
t1; we must consider the variations themselves as functions
of time if we are to properly evaluate the  sensitivity  of the
system to parameteric  variations or uncertainties.

     If the set of coupled  differential equations describing
the system is  non-linear,  not only  the value  of  c.  at  time
t may  in general be altered, but the  shape  of the solution
curve  itself may change.   In certain  extreme  cases, changing
some  parameters  p   (s f j) by small amounts may result in
                         18

-------
               t
                                                                 ^.,..
 x|
 o
Figure 1 (a)
The effect on the solution for c^ as a function of time
when one of the parameters p- is altered by +_ 6p-, where
Pj is one of the rate coefficients or boundary conditions.
( For example, c^ may represent the concentration of ozone
and p. may represent a local photochemical rate coefficient.)
The parameter p. may also represent the initial value of a
different concentration c«,
 x|
      i,I=Pk
              A\
                                                            '"  "c1(x,pj+6pk,t)
Figure 1( b).
The effect on the solution for c- as a function of time
when the parameter p,  representing the initial value
of c.j is altered by  HH 6p. .  Because of the non-linearity
of the system of coupled equations, the solution curves
are not the same shape as the solution for the "best value"
with a constant displacement.  Changing the initial value
can alter the qualitative shape of the concentration-versus-
time curve significantly.
                 19

-------
 changing the solution curve from a smooth,  gradually
 varying function of time into a very  rapidly  oscillating
 curve of small amplitude which follows  generally  the
 trajectory of the1 unperturbed solution,  in  such cases,
 the magnitude of the  variation  6c.].(_x,t)  may  always  be
 small, but we would certainly want to  describe  the  system
 as highly sensitive to variations in  p  .

     Other measures are obtained by introducing  statistical
 averages taken over distributions of  parameter  values.
 For example, mean species concentrations, »  are
                                              I
 defined by

     E /.../ ci(2clp1%,pM; t-)P (P1,...,PM^dpr ..dpM  }

                                          i =  1,N,       (2.8)

where P (PJ >P2 > • • • »PM)  represents  the  probability distribution
for the parameters.  Uncertainties in  species  concentrations
are often expressed in terms of  variances,  a^ ()(_', t'), defined by


    a^x'.f)  E<  [cjCx'.t1) - ]  2>             (2.9a)
or
    Oi(x' ,t')  =  - 2   i  = 1,N,    (2.9b)
where
       .(xI,t')>  = /.../  c.(x';Plj.,.p  ,t')
                                         2
                                            i  =  1, N     (2.10)
Variances  provide  convenient  measures  of expected uncertainties,
which were  discussed  above  in  connection with AIR QUALITY
POLICY  motivations.   As  for expectation  values of sensitivity
coefficients,  sampling   methods  have  been developed to evaluate
measures  of , which will  be described in Section 3.
               I , K
                          •20

-------
GRAPHIC INTERPRETATION

     A graphic interpretation of the preceding discussion
may clarify the meanings of, and differences among, the
various measures of sensitivity and uncertainty.  For
simplicity, suppose that the set of coupled differential
equations (2.1) depends on only two uncertain parameters,
pi and PO» whose values lie between p,    and p,   and
between Poln and p^*, respectively.  For any given position
and time (x,t), the solution c- can be considered to be
                               I
a function of p, and p2 as shown in Figure 2.   The "best
values" of p-^ and p~ are, respectively, p° and p2; in
general, these are not in the exact centers of the respective
domains of uncertainty.
     The point Q on the solution surface represents the
magnitude of the solution c- at time (x.»t) computed with
both uncertain parameters assumed to have their "best values.'
If p, were then varied over its domain of uncertainty,
holding p2 fixed at its "best  value," the solution point
would trace out the curve DQD1 .,  Conversely,  if P2 were
varied over its domain of uncertainty, holding p, fixed at
p°, the solution point would trace the curve AQA1.  Varying
both of the parameters over the full domain of uncertainties
in both then generates a two-dimensional solution surface
for each variable c- for each  position _x.  This solution
surface changes as t itself is allowed to vary.
     In the general case, far  more than two uncertain para-
meters are involved in an air  quality model.  If M parameters
are uncertain, the solutions trace out M-dimensional hyper-
surfaces; the features of the  analysis discussed below  apply
equally well to the M-dimensional case as to the two-
dimensional case.
                          21

-------
                                                       Range of
                                                       uncertainty
                                         Domain of uncertain
                                         in the parameters
Figure 2.   Schematic 'representation of the solution surface
            c- over  the  domain of uncertainty in the parameters.
            Tne resulting  range of uncertainty in the value of
            the solution c^ about the nominal value corresponding
            to the "best values" of the parameters is also
            indicated.
                        22

-------
     The solution surfaces c-({p}) may have very complicated
topographies, with peaks, pits, undulations, sharp ridges
or gorges, saddles, and even such discontinuities as sheer
cliffs.   The purpose of sensitivity analysis, from this
point of view, is to understand the shape of the solution
surfaces as we vary the uncertain parameters away from the
immediate vicinity of the current "best values."  As new
field data or new experimental  evaluations of parameters
become available, our estimation of the "best values" may
change,  moving us across the complex topography of the
solution surface.  Understanding the shapes of the solution
surfaces provides advance warning about what kinds of changes
in the parameters will most drastically change the predictions
of our simulation models.
     As  we mentioned earlier, for policy-making purposes
uncertainty analysis is generally of greater interest than
sensitivity analysis.  The purpose of uncertainty analysis
is to assay the range of uncertainty in the solutions c-
resulting from the present degree of uncertainty in the
parameters, as indicated in Figure 2.  A crude measure of
this uncertainty is the difference in elevations between the
nominal  case (point Q) and the extremes (points B and C').
Although the extreme values of c^ are shown in the figure
to lie at corners of the domain of uncertainty in the para-
meters,  the extreme values will more commonly lie somewhere
inside that domain.
     If we have some idea about the probability distributions
for the  uncertain parameters, we can in principle compute
the probability distribution for c. over the domain of uncer-
tainty in the parameters, as shown in Figure 3.  From the
probability distribution for c^ , quantities such as mean
species  concentrations  and variances a.= can be computed
                          23

-------
           Ci(x,{p},t)
                              D1   C1 max
                               	 -c-r- -
Figure 3.   Given assumed probability distributions for
            each uncertain parameter, the probability
            distribution for the solution c- is obtained
            by computing a probability-weighted distribution
            of elevations of the solution surface. Note
            that the "best value" of a parameter may differ,
            in general, from both the most likely value
            and the mean value of the parameter. Likewise,
            > the mean value of the solution c^,
            may differ from the most likely value of the
            solution and from the nominal value c.(Q)
            corresponding to the "best values" of the
            parameters.  We can never be absolutely certain
            that the true value of a parameter lies within
            a given domain of uncertainty.  Thus P(p-i) and
            P(p«), in general, have small but finite values
            beyond the domain shown.  In this illustration,
            the uncertainties in p, and p,-, are assumed to
            be independent and uncorrelated.
                    24

-------
by standard statistical methods as defined in Eq's (2.8),
2.9) ,  and (2.10).
     Several important points should be noted in Figure 3.
First, the investigator's estimate of the "best
value" of a parameter may or may not coincide with either
the most likely value or the expected (mean) value of the
parameter as computed from the probability distribution.
(Note, in particular, p, in the figure.)  Second, the shape
of the solution surface c- is independent of all assumptions
about the shapes of the probability distributions for the
parameters; the shape of the solution surface depends only
on the structure of the set of coupled differential equations
(2.1).  Third, the probability distribution for the value
of the solution c^ (.x.t) depends on both the assumed probability
distributions  for the parameters and the shape of the solution
surface.  Thus the mean value  may or may not coincide
with either the most likely value of c^ or the nominal value
of c-  corresponding to the "best values" of the parameters.
Fourth, the variance o. of the probability distribution for
the solution c- will in most cases be significantly smaller
than the range of uncertainty in GI , that is, 2ai<_(c!jiax-crjn n).
     The significance of Eqs. (2.3), (2.4), and (2.5) can be
seen in Figure 4.   If we make small changes in the para-
meters, one at a time, and compute c- for the altered para-
meters, we  find in general that c- has changed by a small
amount Sc-.  Changing Pj by a small amount 6pj moves the
solution point from Q along the curve DQD1 toward D.  (See
figure.)  Changing p2 by a small amount 6p2 moves the solution
point from  Q along the curve AQA1 toward A1.  The differences
in elevation resulting from these small displacements within
the domain  of uncertainty are given by Eq. (.2.3).
                          25

-------
                                                   rain
                                         max
Figure 4.    Construction of first-order sensitivity
            measures and coefficients.  First-order
            sensitivity analysis approximates the
            true solution surface bounded by the curves
            BDB',  BAG,  B'A'C',  and CD'C' by the tangent
            plane at point Q which is bounded by the
            straight line segments 363', Bay, B'a'y',
            and Y^'Y'J  respectively.  Using this approx-
            imation severely underestimates the range
            of uncertainty in the solution c. and provides
            no warning of such structural features as the
            deep valley at V.
                    26

-------
     In the limit of infinitesimal displacements Sp^,
the ratio of the deviation in c- to the displacement in p.
is the slope of the solution surface in the direction of the
p. axis at the point Q.  The slopes thus defined are the
 J
first-order sensitivity coefficients Z- .  defined by Eq.
                                       • »J
(2.5).  First-order sensitivity analysis,  in effect,
approximates the shape of the solution surface bounded by
the curves BOB1, BAG, B'A'Ci and CD'C1 in  Figures 2, 3,
and 4 by the plane bounded by the straight lines 36g',
BCCY'J B'a'y'' and Y^'Y' which is tangent to the solution
surface at the nominal solution Q.  (See Figure 4.)
     For very small displacements about the "best values,"
the true solution surface differs from the tangent plane
by only small differences in elevation.  In this regime,
comparison of first-order sensitivity coefficients tells
us to which parameters the solutions are most sensitive.
But first-order sensitivity analysis provides no warning
whatever about such significant features as the deep valley
at V, only a modest distance in parameter space from the
"best values."  Nor would this analysis indicate the full
range of uncertainty in the solutions c- over the domain
of uncertainty of the parameters p .  For the case shown
in Figure 4, the range between extreme values on the tan-
gent plane--  [C-J(Y) - c-j (31)] -- is less  than half the
range  for the true solution surface--  fc.(C') - c.j(B)j  .
Moreover, the values  of the parameters corresponding to the
extreme values on the tangent plane are completely different
from the values  of the parameters for the  true extremes of c-
                                                            I •
      If we wish to learn something about the shape  of  the
true solution surface at finite distances from Q, we can
perform a Taylor series expansion about c^Q), evaluating
not only the first derivatives of ci with respect to all
the parameters p., but also higher-order derivatives
                J
(higher-order sensitivity coefficients).  Including second-
                          27

-------
order coefficients  would be an improvement over first-
order sensitivity  analysis, but it would be inadequate
for the case shown  in Figure  4 since Q  is an inflection
point of the curve  DVQD1.   Thus the second derivative
 2     2
9 c-/9pi vanishes at Q,  and we would still have no intima-
tion of the valley  at V.
     In general, a  truncated Taylor series approximation to
the solution surface can be characterized by a radius of
approximate convergence , a small  distance in parameter space
about the point Q within which the series approximation differs
from the true solution by less than some specified e.  Regard-
less of how many terms are included in  the expansion, however,
the difference between the approximation and the true function
depends on the still higher-order terms which have been omitted,
Unless we have sound theoretical  reasons to show that higher-
order derivatives become vanishingly small (and this  is almost
never the case for realistic  air quality models), we  have no
assurance that we will not overlook such features as  the
valley at V    a very short distance away from Q.  Consider,
as  an example, a function whose first and second derivatives
are of order unity, while all higher derivatives vanish,
except the 20tn, which has a  value of 10   .  The terms
involving this derivative are then of order (10   )  <5p  ,
                            - 5
so  that to achieve an e ~ 10   , the radius of convergence
would still be $10~  , even if we included all terms  up through
the 19th order.
     An alternative  approach  is to systematically examine the
solution surface by  computing c. from the simulation  model  at
a  significant  number of locations  in the  domain of uncertainty
of  the  parameters  {p}.  Unless the functional forms  involved
 in  the  various coupling,  source, and sink terms in the set  of
differential equations  (2.1)  are simple expressions,  it is
difficult  to be  sure we have  not accidentally overlooked
                             28

-------
significant structures hidden between the locations actually
examined.   Several  of the sampling methods discussed in a
later section attempt to deal with this problem in one way
or another.
     We noted earlier that the shapes of the solution surfaces
c.j ({p}) depend only on the structure of the basic set of
coupled differential equations (2.1).  If the functional
forms for the sources, sinks, photochemical  rates, or chemical
kinetics rates are  changed, or if new reactions are added to
the system, or other reactions are removed from the system,
all of the solution surfaces are changed.  Peaks, pits,
undulations, ridges, gorges, and cliffs in the surfaces
corresponding to any particular chemical species c., for
example, will all  be altered; some may disappear, while new
topographic features may appear.  The shapes of the distribu-
tion functions P(c.) deduced for the values  of the solutions
c- (see Figure 3)  and the range of uncertainty for c- (see
Figure 2) will also be changed.  Thus, whatever method of
sensitivity analysis or uncertainty analysis is used, the
evaluation of sensitivity or uncertainty must be repeated
ab i n i t i o each time the system of equations  (2.1) is altered.
                            29

-------
                        SECTION  3
         SYSTEMATIC SENSITIVITY  ANALYSIS  METHODS

    Concise descriptions are given in this section of the
candidate systematic sensitivity analysis methods, starting
with the logically simplest "time-honored" method and
proceeding next to variational methods and, finally, to
sampling methods.  We subdivide this section into two
major subsections:  Deterministic Methods and Sampling
Methods.  The deterministic methods include both the "time-
honored" and the variational methods.

    Arbitrary, unaveraged deviations, , are the
                              I             IK
measures used in   sampling  analyses.  Table 1 summarizes
the salient features which appear in this section.

DETERMINISTIC METHODS

The Time-Honored Method

    Trial-and-error testing of model parameters has stood
the test of time in all scientific disciplines.  This time-
honored method continues to be the most widely used sensitivity
testing method because  it is  flexibly applied to arbitrary
problem applications and, most significantly,  because
most researchers understand the trial-and-error method and
are comfortable with it.  Developments in our present study
suggest that the time-honored method will continue  to have
an  important role  to play in  conjunction  with applications
                          30

-------
of more-highly automated sensitivity tests to complex air
quality simulation models.
    The time-honored sensitivity method is applied by
    ily altering the value of model par
(x. ,t) and then noting the deviations,
simply altering the value of model parameter, p., at any
                                               J
    6Ci(x ,{p},t) = c.j(x ,p9 + 6p.j,t)- CjCx  ,Pj,t),  (3.1)

which are obtained from the respective model solutions
for c.j(p° + SPJ) andc-j(pp. (see Figure 4.)  The major
features of this method are:
       Arbitrary deviations 6ci are economically  obtained  for
       applications having poorly defined probability
       distributions- of model  parameters.
       Variances a,(x't') and expectations of sensitivity
                               (2)
       coefficients < Z • .>, < Z. '.  >, etc., can be evaluated
                       ' » K      1 , J K
       but may be extremely costly to compute.  (Examples
       will be given in Section 4.)
       Sensitivities and uncertainties which are  continuous
       in time, t, are  obtained for the entire  problem  time
       domai n.
       It is  a non-dedicated method; that is, the  time-
       honored method  applies  to arbitrary air  quality
       simulation models without any need to alter the  air
       quality simulation model itself, or to develop special
       programming for  performing sensitivity analyses.
                         31

-------
    •  Variations in photochemical  kinetics  and  fluid
      dynamics parameters  are  treated  with  equal  facility
      (whether the chemical  kinetics  and  the fluid dynamics
      are treated separately  or in fully coupled  form)
      over the entire space and time domains  of interest.
      Applications to date  include the full range of  coupled
      kinetics and hydrodynamics.

    Clearly, the time-honored  method has a  tremendous
combination of simultaneous features which  other  sensitivity
methods do not yet have; and,  depending  upon  the  craft
and specific motivations of the practicing  scientist,  the
method is also both systematic and economical.

Variational Methods
      As we  discussed  in Section 2 above, the basic task of
sensitivity  analysis  is to understand the  shapes of the
solution  surfaces  defined  by considering the solutions
c-(x, (p},t)  at  any  given  position  and  time to be functions
  I
of  the uncertain  parameters of the  system.   If we consider
such  a solution  surface (as was illustrated in Figures 2,
3,  and 4)  at a sequence of times, the surface  itself
changes  shape with time.   If we are interested in just a small
neighborhood of the  "best  values" of  the parameters (p>,
we  could  in principal  compute  the solutions c^C^.t) for a
large number of sets  of values of {p}in  the vicinity of the
"best values" (p0) and then laboriously  construct the  solution
surfaces  for each time of  interest  to obtain numerical
                                                       (2)
approximations to the sensitivity  coefficients Zi k, Zj ^  ,
etc., by straightforward  but  tedious  means.
                          32

-------
   But a far simpler method is to treat the sensitivity
coefficients themselves as dynamic variables.  In other
words, we want to obtain equations of motion describing
the change in time of the slopes and curvatures, etc.,
of the solution surfaces.  These can be derived analyti-
cally and then solved once by numerical methods instead
of solving the original set of equations (2.1) many, many
times for different sets of values of {p} in the vicinity
of {p°}.  Variational methods, in brief, are based on
this  latter approach.

    (i)   Chemical  Kinetics with Constant Rate Coefficients

    Most automated sensitivity analyses implemented thus far
have been applied to the case of purely chemical kinetics in
which rate coefficients are held constant in time.   Variational
methods  are also sometimes  referred to as direct methods
when problem parameters are held constant.   By any terminology,
variational methods are computationally dedicated in the sense
that a set of equations for sensitivity coefficients must
be derived and solved in conjunction with the original model
equati ons.

    Photochemical  kinetics model equations  are nonlinearly
coupled ordinary differential equations of the form

    ti = dc./dt = RjU.-.fp}, t),      i,j
                          33

-------
In the absence of spatial dependences, the parameter array
{p>  contains all of the variables listed in Eq.(2.2),
except c.j B  , _y_  , and KD-

    The sensitivity coefficients are defined as'  partial
derivatives of the solutions cn- with respect to the uncertain
parameters:

    Zi   ~-    - '       i=1»N>  k=l,M.           (2.5)
Dynamic equations of motion for these coefficients can
be obtained by differentiating (2.5) with respect to  the
time.  Assuming that the order of differentiations may
be interchanged, we obtain
                               3R:     N   3R.  7
                                  '   •  *"=   	L  ^i k
                                                 J » K   »
      i,k    dtu.  r ,D   v"i;    ,D      L
                    3pk         9pk     j = l
                 i,j=l,N,    k=l,M.                (3.3)

 Considering  the  solutions  c-  as  elements  of  a  column
 vector  of  length  N, we  can  think  of  the arrays  Z. , r,   Z-,u   and
  i                                               1  K. "   1  K
 R.  ,  E  3R./3p. either as matrices  of N by  M  elements  or  as  sets
  1 5 K      i    K.
 of  M  column  vectors each of length N.  The Jacobian matrix
 J.  .  =  3R-/8C- is  a square  matrix  of N by  N  elements.   If
  ' 9 0      '    \J
 we  view the  set  of N by M  equations  as M  vector equations,
 we  can  write

                 + J- (Z  I .  ,       k  = 1,M,    (3.4)
               k

 where it is  evident  that the  columns  of Z.  .  are  completely
                                          I , K
 uncoupled from each  other.
                          34

-------
    Important features of this type of sensitivity analysis
application are:

    •  Eqs. (3.3) can be automatically compiled and solved           •
       for arbitrary chemical  mechanisms in the same manner
       as the chemical kinetics  equations (3.2)^ '.  This
       eliminates tedious and  error-prone hand-programming
       and thus allows very rapid turnaround  for solving             !
       arbitrary  chemical mechanisms, in both the kinetics
       and the sensitivity programs.                                  ,
                                                                     I
    •   Although the  M vector equations  (3.4)  are decoupled           <
       from one another,  both  R1  and  J  depend explicitly on
       the current  values of c^  which are determined by
       solving (3.2).   To obtain  fully  reliable solutions,
       each of the  M equations (3.4)  should be solved
       simultaneously with (3.2),  resulting in a system  of
       2N simultaneous differential  equations to be solved
       to obtain  each of the M column vectors making up  Z. ^.
                                                         I , N
    •  The cost of simultaneous solutions, using modern
                          / o \
       stiff  ODE  solvers,^ 'for the  combined  set of kinetics
       and sensitivity equations greatly exceeds the cost
       of solving the kinetics system separately from the
       sensitivity equation system.  It appears possible
       to  reliably accomplish  this separation  in many
       photochemical  problems, although some  additional
       research is still needed on this aspect  of  the
       problem.
    •  The method is  first order.  Although  equations can
       can be similarly derived for  higher order sensitivity
       coefficients,  ZJ^, ZJjkr etc.,  no consensus has yet
       emerged on the value which would be derived  from
       higher  order  evaluations.  (For many  research users,
       first  order terms  provide sufficient  information, whereas
       for some policy applications  the contributions of             ;
       higher-order  terms may also be desirable. (See Eq.(2.4).)
                         35

-------
    •   Sensitivity coefficients  are  evaluated continuously
       over the entire  time  domain  of interest.
    t   When statistically significant information about
       the  distribution of input parameters  is available,
       output  variances, valid  to  first-order,  can be
                 (9 )
       evaluated.v '

    (i i)   Chemical Kinetics  with Time-Varying Rate Coefficients

    When  rate coefficients are varying in time (e.g.,
diurnally varying photochemical  rates) the sensitivity
coefficient is defined as

     Z = O/3e)  fc-Uj, ... ,CN;  . PR( t )+egk (t) ;t]       ,  (3.5)
                                                 e = 0
where the  function gk(t)  measures the uncertainty  in  pk(t).
The sensitivity equations are then given  in analogy to Eq.
(3.4), by
     Z = R(t)  + J- Z.                                    (3.6)

The functional  derivatives  Rk(t) can  be  evaluated,  for the
kth parameter,  in  variational calculus,  by
      Rk(t)  =  (3/3e) [R(ca,..,cN; pk(t)  + egk(t); t]
e=0
    (3.7)
      This  method  has  been  applied  for  diurnal  variations  in
 a  simplified  atmospheric  ozone  example  by  Dickinson  and
 Gelinas^  ,  and specific  choices  for  the  function  9i<(t)
 have  been  indicated.   The  important features  of the
 diurnally  varying examples  were qualitatively  much the same
 as  in cons.tant  rate  coefficient case  which was discussed
 above,  but with some  operational  distinctions.

      •   The evaluation  of  functional  derivatives  requires
         some  dedicated  programming of  the  functions  gk(t)
         in  the  sensitivity  program.
                         36

-------
     0  Kinetics and sensitivity solutions can be much more
        costly to obtain when rate coeffievents are varying
        with time than  when alternative constant rate
        coefficients are used.  For example, experience  in
        diurnal phothchemistry has shown^  ' that diurnal
        kinetics integrations can be ten to fifty times
        more costly than corresponding constant (daily
        averaged) rate photochemistry.   This proves to  be
        a serious practical constraint for all dedicated  and
        non-dedicated sensitivity methods.

     (i i i)   Non-Reactive Hydrodynamics
     Examples of variational sensitivity  methods,  as well
 as  the Fourier Amplitude Sensitivity Test, which is  a
 sampling method, have recently been presented and compared
 by  Koda et  al^   , they  present several computational examples
 for  the atmospheric diffusion of inert species while also
 discussing  reactive systems.  The governing non-reactive
 hydrodynamics  equations are  (in one spatial dimension).
     3c.(x,t)   8    r      3c,(x,t)  .
     —	 = —   [KD(x) —!-	 1   .      i  =  i.N,
                8x                    J
where KD(X)represents spatially varying diffusion coefficients,
which are assumed to be identical  for all  species.  Boundary
conditions are represented by

     - KD(0) —- = S.(t)       at x = 0,        (3.8)
 where S^ represents all sources  and sinks of the i-th chemical
 species  at time t at the boundary x = 0,  and
     Hi.
     3x  =  °                   at x = H(t),   (3.9)

where H(t) is a time varying boundary height.  Initial
conditions, c- T(x)  = c-(x, 0), are  usually assumed to  be
              1,1-       i —
known with no error.
                         37

-------
     Sensitivity equations are obtained by first discretizing
Eq.  (3.7) in the spatial domain, such as by the method of
lines.  ^  The partial differential equations (3.7) are  there-
by reduced to a large, finite set of ordinary differential
equations in which c.(x,t) and KD(X) are evaluated'only  at
discrete  spatial locations  x£, £ = 1 ,L.  For a single
species the discretized form of Eq. (3.7) is (in vector
notation, and including boundary conditions)

     dC(t)
     —  = A(Kn)-C(t) + S(t)5                    (3.10)
     dt         U
where C(t) = [c(Xpt), C(x2,t) 	C(xL,t) ] T, KQ = [ KD  (xj),
KD(x2),. . . ,KD(xL) ] T, and S(t) =  |s (Xj ,t), 0 ,. . ,oj T.  Initial
conditions are denoted by CQ = C(o)» and the matrix A(KD) is
a finite  difference representation of the diffusion terms in
Eq.  (3-7).  For constant KD, the analytic solution for Eq.
(3.10)   is
                            t
     C(t) = e A(KD}t  [c  +  C  e"A(KD)1:l S(t')dt']  .  (3.11)
                      L °    0
                            o
Sensitivity coefficients for this case are  defined by
                                                    T
              3C(x,,t  )    3C(x9,t)        3C(x, »
     Z,(t) =      l           2     	     L
      O
                          9KDj
                                           j=l,L.        (3.12)
 where KD-  sKD(x-).  A sensitivity equation (in vector form) is
 obtained  by  a  derivation analogous to that used in the develop-
 ment of Eqs.  (3.3)  and (3.4).   This development yields the
 sensitivity  equations
dZ-(t)                     3A(K)
                                 .  C(t),  j  =  1,L,
                                                  (3.13)
              =  A(KJ  '  Z,(t)  +  - ^- .  C(t),
                         J
 which  can  also  be  solved  analytically.
                         38

-------
     Important features of this method are similar to those
mentioned in (i)  above, namely:
     •   The  solutions  to Eq.  (3.10)  and (3.13)  can be
        evaluated either analytically or numerically
        for  arbitrary  applications.
     •   The  diffusion  coefficient at each spatial  grid
        point is  treated as an individual, uncertain
        parameter.
     •   The  method is  first order.   Higher order terms
        have not  yet been evaluated  for this  case, to
        the  best  of our knowledge.   (The utility of higher-
        order terms is application-dependent.)
     •   Sensitivity coefficients are evaluated  continuously
        in time over the entire space and time  domain of
        i nterest.
     •   This method can be generalized to also  include (constant)
        wind advection terms.   In such cases,  the wind values
        assigned to each grid  point  are also  treated as individual,
        uncertain parameters.
     •   Output variances, valid to first order, can be
        evaluated when statistically significant information
        about the distribution of diffusion data is available.

     (i v)  Reactive Hydrodynamics with Constant Rate
           Coefficients and/or Diffusion Coefficients
     Sensitivity  coefficients  are defined for the i-th species
(i n vector form)  by

     Zi,j *k.C1(t> - IP,
              J          J
                         1=1,N,    j=l,(M+L),      (3-14)
and the spatially discretized  sensitivity equations are given
by:
3Z-  .(t)
  1 »J     _
3t
                3A(KD)
.  +

                                     A(Kn) + J
                    1=1, L,     j=l,(M+L).
Z,  .(t),
                                                   (3.15)
                          39

-------
In this case,  p.  spans all  chemical rates k,...kM and
diffusion coefficients KD ^ -,..., KD L at all  spatial  grid
points.  The total number of parameters is (M + L).

     To our knowledge, applications at this level of complexity
have not yet been done for air quality simulation models.
The important  features follow the same general lines as
in the cases which have previously been discussed.  Clearly,
the amount of  dedicated programming and data management can
approach   prodigious proportions for this level of applications
where there can easily be 30 chemical species, 75 photochemical
reactions, and hundreds of grid points.

     (v)  Reactive Hydrodynamics with Space and Time Varying
          Coeffi cients
     Cases in  which Kn(x), S.(t), H(t), and k- can vary
                       "          (n)        J
have been considered by Koda et al^1.  Equations have been
developed for the functional derivatives,  6C - (.x1 , t' )/6 K[)(x) •
6ci (x1 ,tJ)/£S.j(t), &c.(x' ,t')/6k..,  and  &c- (x', t' )/
                   j A   j             j i   j
                                                    (3.16)
                          40

-------
subject to initial  conditions
     6ci (x, 0)  = 0,                           (3.17)
and boundary conditions
            36Ci(x.,t)         3c?(x»t)
     -K°(0) -—	— - 6KD(0)—	  = 6S1(t) at x = 0 ,
            3x                3x             (3.18)
and
     36c,(x,t)
        J
     3x
= 0         at x = H,  i=l,N.    (3.19)
          suggested that  input parameters be  expressed
as expansions of orthogonal  basis functions,  which  leads
to the following development of sensitivity equations:

     (a)   Select a set of functions {<];•  (x_,t)}.
     (b)   Multiply Eq. (3.16) by \l>- and sum over  i  =  1,K.
     (c)   Integrate the resultant equation in (b)  over  the
          intervals ^x = [o,HJ  and t = fo,fj .
     (d)   Apply the initial and boundary  conditions,  Eqs.
          (3.17) - (3.19).
     (e)   Specify a defining equation for ij>.(x,t)
          3^.    ~        Sty.       "n°
          7T " ^ (K°D(-)"97')"
          wi th
          ax
            i= o, x = [O,H].                        (3.21)
                          41

-------
     (f)  Specify a terminal condition  on  the { ty.},  e.g.,

     ^(x.T) = 6ijjS( x -  x') ,    1.j = l,N,

                                                (3.22)



where 6.. and 6(x - x1) are, respectively,  the  Kronecker
        ' J       ~   ~
and Dirac delta functions.
The results of performing operations  (b)  and  (c)  and  solving

Eq. (3.20) gives   N   H         J    q

     Sc.(x',T) = - I   S S*D (^  S  at1  Iw^1 dt dX
       1 —             n         r\  0 X    OX
       N    T

     + E    S  6S..(t) ^j^O,

       .j=l  0

       M      N   H  T  aDo
                  ?s  r,.  o K ^
          6k,
              r-   .»>  r> on--i
            J I   -J  i     ^*i dt dx-           <3-23)
               =100 8kj
The functional derivatives which we  seek  are


                 N     J   °
               =  r^    \
     6c-(x1,T)
       '          r^    \    J   J
     6Kn(x)           n 9x
       D ~       j=l  U


     6c,(x',T)
     — - -  =
                 il'j^O.t)   ,                     (3.25)




                N    H  T   3Ro

                        S   —— ^o-i dt dx  .     (3.26)
     6kj        £=1  0  0  3kj
                          42

-------
     A similar exercise can be carried  out  in  order  to
obtain Sc • (x ' , t ' )/6H ( t) when H is a  function of  time.
          1
Sensitivity  functions  (3.24) to  (3.26)  are  obtained  by
solving a decomposed equation for the nominal  values
c?(_x,t),  and solving Eq,  (3.20)  for  ^£j(x_,t) in  order
to carry  out the operations in (3 . 24)- (3 . 26) f   Koda
et al ^   ' have obtained  solutions for

     <5c(x't)   6c(x' ,T)     ,  6c(x',T)
     _ — _             , d 1 1 U  _ ~ __
     <5KD(x)   '  6S(t)           6H(t)

for a single non-reacting species in  a 1-D  (vertical)
boundary layer simulation and  have related  the resulting
functional derivatives  at specific points to sensitivity
functions obtained in  a discretized grid calculation which
was discussed in  (iii)  above.   The relationship at spatial
location x.'  between sensitivity coefficients which are
developed from functional derivatives and those which are
developed from discretized grid point formulations i's
                             x
                'a £,• ( x ' , t )    ^      6 d . ( x ' t )
      -.(x',t) E—L-Z -  =  \         1 ~
                                                      (3.27)
     ac,(x',t)      6c,(x',t)
     _J -- -  = AX - ^— = -
                                                     (3.28)
where Ax represents an appropriate grid spacing and 6p,-(x)
                                                      J
is assumed to be zero at grid points other than x-
                         43

-------
The major features to note for the variational methods
involving functional  derivatives are:

     t  The solution  of the system equations for nominal
        values, <:•(>(,t), and of the adjoint equations for
        tyo-(x.\t) are  complex calculations, even with no
         JC 1
        reactions present; but the method produces highly
        rigorous sensitivity measures.
     •  This highly dedicated method will be difficult  to
        automate sufficiently for reliable, rapid turnaround
        calculations  in arbitrary model applications.
     •  The method is first order and subject to the same
        considerations which were previously discussed.  No
        higher  order evaluations have yet been made.
     •  Sensitivity coefficients are evaluated at specific
        space  and time points (xjt1) but  rigorously  contain
        correlative effects which have occurred along
        characteristic trajectories which lead to (_x',t').
     •  Arbitrary deviations are the measure of sensitivity
        in  this method.   It is not yet apparent that
        expectation values of observables could be readily
        obtained when meaningful parameter  statistics are
        available.

 The Green's  Function Method
     Hwang, Dougherty, Rabitz, and Rabitz^  'have developed
 a variant  of  the direct variational method of sensitivity
 analysis for  chemical kinetics systems.   The basic  develop-
 ment of the nethod is similar to that described in  our
 discussion of  the variational method for chemical kinetics
 with constant  rate  coefficients.
     Starting  from the kinetics model equations (3.2),  the
 dynamic equations for the evolution of the sensitivity
                          44

-------
coefficients Z^ k are derived as before, obtaining equations
(3.3) or (3.4).  As we remarked in the earlier discussion,
Eq. (3.4) can be viewed as M uncoupled equations, each des-
cribing the evolution of a column vector of length N:

    (Z)k =(R')k + J'(z)k '             k = ^M.     (.3.4)

In the most general applications of the direct method,
each one of these equations   is solved numerically together
with the basic kinetics equations (3.2) in order to allow
the stiff equations system solvers to adjust the time steps
optimally for the numerical  integration of each of the vector
equations in (3.4), without introducing numerical errors due
to interpolation from an independent solution run for the
basic kinetics equations (3.2).
    The Green's function method continues its development
by assuming that each of the vector equations in (3.4) can
be reliably and safely solved by itself, interpolating
values of c. obtained by a careful separate integration of
(3.2).  Since Eq. (3.4) is linear in Z, the solution of
the i nhomogeneous equation can.be expressed in terms of a
Green's function K.j^(t,T) as
                                Tl R;R UI,...CN; {p};T )dt,
                      0  i=l
                               i,£ = l,N,  k = l,M,   C3.29a)
if Pk is one of the initial conditions, where R'. = 8R. ,3p  ;
or as
N
E
R=l
    ZiJ<(t) =        K(t,r) R   (c ..... C; {p};r) dt
                                                 (3.29b)
if pk is  any other parameter of the system.  In either case,
                       45

-------
the Green's function is the solution of the homogeneous
equation
               N
    Kik(t,T) - JT   Ju (c15..., CN; (p};t) Ku(t,T) = 0,
               1=1          i,*5k = 1,N,        (3-30)
where  J ^  E 3^.7 3cA  is  the Jacobian  matrix as before, subject
to the causality condition
    I  (t,r) =  0   for  all  t N),  which  is usually   the  case  for realistic  simulation
model s .
    Practical  implementation  of the Green's  function method
proceeds  as follows:
     (a)   The  simulation  model  equations (3.2)  are  integrated
numerically using  the  nominal values 'j> } f or the uncertain
parameters  for times  t=0  to t=T,  recording or storing  values
for the solutions  c^  at  numerous  intermediate points in time,
     (b)   The  time  interval T is subdivided by a coarse arid  of
1 + 1 quadrature times  xm,  TQ= 0 
-------
    (c)   Eq.  (3.30)  is solved numerically for each  of the
subi nterval s  x  ,< t< T  with initial  conditions
              ffl — J.      HI
K-uU  .,  Tm  ,) = 6...  The values of  c-(t)  needed  to evaluate
 i K  m -r I    in — i     IK                  i
J,o(t) are interpolated as needed from the stored solutions.
    (d)   The  sensitivity coefficients  themselves  are computed
by using  the  appropriate form of Eq.  (3.29), replacing the
integrations  over time by numerical quadratures on  the coarse
 grid  T .   It can  be  shown  that  the Green's  function
for any  two points on the mesh may be  obtained as a
product  of the Green's functions obtained for the inter-
vening subi nterval s :
                  N        N
                             m,nr,        (3_33)
     (e)  If higher order sensitivity coefficients are
desired as well, they may also be computed by quadratures
with the same Green's function:

                              vt>T)
                                          (3.34a)

where the differential operator D- is defined as
                  N              ^
           3    + T   7      ^
     Dj =  -    L   Zq,j  —       •    (3.34b)
           ^p,-    q=i       acn
             j                q
     The method is clearly a dedicated method, requiring
new  programming for each  problem, but it is highly amenable
to automatic programming  methods just as the chemical
kinetics equations are.
                        47

-------
SAMPLING METHODS
Monte Carlo Method


     Stolarski et al^ '  have recently presented an atmospheric

uncertainty analysis which is based upon Monte Carlo selection

of trial values for chemical rates in a 1-D stratospheric

model which gives steady state solutions of Eqs. (2.1) with

v^ = 0.  Many of the photochemical reactions of stratospheric

interest have been measured often enough and well enough

that it is reasonable to assign probability distributions for

the parameters describing these reactions.  Most rates which

are known to be significant in the stratosphere are uncertain

within  factors ranging from 1.1 to 10.

     Hypothetical reactions must simply be included and

assigned nominal  values and uncertainties on a tentative basis,

The Monte Carlo method is executed as follows  (see Figure 5):

     (a)  A log-normal distribution is assigned to the
          probability that true reaction rates lie between
          (k°j/r.) and rjkj°, where r->l is the factor by which
          k-j isjuncertain.          •'
           J

     (b)  The atmospheric model is run, using  a chosen "best
          set" of reaction rate values and center line values
          for other parameters.

     (c)  A random number generator is used to select trial
          values for the entire set of reaction rates, simul-
          taneously.

     (d)  The atmospheric model is run again,  using the set
          of simultaneously perturbed trial reaction rates
          from (c).

     (e)  The processes (c)  and  (d) are repeated until enough
          output samples are accumulated  to evaluate
          statistically meaningful mean values and variances
          of the output variables.  Stolarski  et aZ(l5) ran
          sufficient cases so  that the accumulated 10 im-
          precisions on both sides of the output distributions
          converged to within  1 or 2%.


                               48

-------
Figure 5.   Schematic illustration of the Monte Carlo
            method of sensitivity analysis. A random
            number generator is used to select values
            of all the uncertain parameters within the
            domain of uncertainty. The simulation model
            is then run using those values for the para-
            meters. Values thus computed for c^(x,{p},t)
            are then tallied and analyzed by standard
            statistical methods. The distribution of
            values obtained in this way is illustrated
            above as a histogram, with mean value
            and standard deviations a~*~ and a". Stolarski
            et al assume log-normal distributions for
            the uncertain parameters, but any distribution
            function could be used,  including a joint
            probability distribution for two or more
            correlated parameters.
                   49

-------
     The important features  of the Monte  Carlo method

are:


     •  The method is non-dedicated and thus readily applies
        to arbitrary air quality models.   To date only
        chemical rate uncertainties and boundary condition
        uncertainties have been considered.

     •  The method has, so far, been applied  as an
        uncertainty analysis method which evaluates output
        variances.  Questions of how a shift in one parameter
        value would change the model response to all other
        parameters have not yet been indicated.


     •  Dependences of the method upon selected probability
        distribution functions have not yet  been indicated.

     •  Best and worst case  estimates  can be  identified  and
        retained from the accumulated test cases.

     •  Initial studies are  underway forcorrelated parameters.

     •  Computer economics presently require that the 1-D
        stratospheric model  be physically simplified and
        solved  for steady-state solutions rather than the
        far more costly transient solutions.  Stolarski
        et al emphasize that the feasibility of using the
        Monte Carlo method depends most heavily upon reducing
        the computational cost of the atmospheric model  to
        an absolute minimum.  In their case, complete repro-
        gramming and physical simplifications were required
        for their original,  more general  atmospheric model;
        this went far beyond simply altering the original
        model and obtaining  nominally improved  economies.

     •  Sensitivities were determined  separately by the
        time-honored method  applied to individual ratesr
        The rates were  varied,  one  at  a  time,  by +_ la  and  by+2a.

     •  The number of test cases which were required to  obtain
        statistical convergence in a-j(x^,t)  was proportional,
        in some sense,  to the number of sensitive parameters
        rather  than to  the total number ofuncertain
        parameters.  Convergence was more rapid for  those  species
        and locations which  were most  affected by  rates  with
        small imprecision; relatively  poor  convergence was
        experienced for species such as the methane oxidation
        chain,  for which important reaction  rates are quite
        uncertai n.

                           50

-------
     •  Stolarski et  al     emphasized that the uncertainty
        analysis must be  redetermined when:(i) a new reaction
        is included  in the  model  (ii) substantial changes
        occur in parameter  values  or their imprecisions,
        or (iii) substantial  changes occur in the basic
        understanding of  atmospheric processes.

FOURIER AMPLITUDE SENSITIVITY  TEST  METHOD
     The Fourier Amplitude  Sensitivity Test was first developed
by  Cukier   and Shuler ejt a_]_      to provide a means for
determining the sensitivity of the  solutions c^Cx., (p},t) of
a large set of coupled rate equations to uncertainties  in
each parameter p when averaged (in  a certain sense) over
uncertainties in all  the  other parameters.
     Figure 6(a) i11 ustrates the basic idea of the FAST  method.
For simplicity in illustration, consider a system of coupled
rate equations depending  on only two parameters,pi and  p^.
Although the investigator can select some "best value"  for
each of these parameters, the true  value lies within some
range of uncertainty  about  that "best value".
     If only one parameter  were uncertain, we could profitably
ask, "What is the slope  of  the solution surface as that single
parameter is varied  over  its  range  of uncertainty?"  (See
line AA1 in  Figure  6(a).) But other parameters are also un-:
certain; thus in the  general  case  we do not know where  on the
solution surface we  are.   It is thus of some value to ask,
"What is the average  slope  of the  solution surface in the
direction of parameter p- when we  average over the full range
                         \J
of uncertainties in  all  other parameters as well?"  In  other
words, what is  the  average  slope of curves AA', BB', CC1,
etc.?*
*
 In some cases,  the average slope may be zero or very small for parameters
which in fact are sensitive.  The average slope transverse to a symmetrical
trough or ridge, for example, vanishes, even if the trough is very deep
(or the ridge very high).  To avoid being mislead by such cases, a
squared norm is  sometimes used to indicate averaged sensitivities.
                             51

-------
                       Ci(x,{p},t)
                                                         >P2
                Schematic  illustration of  the  Fourier  Amplitude
                Sensitivity  Test (FAST). The FAST  algorithm
                generates  a  closed search  curve  across the
                domain  of  uncertainty of the parameters which
                begins  at  the  "best values" of all parameters.
                Values  of  c.^ at  each sampling  point (marked by
                large dots in  the figure)  are  tabulated against
                the  search parameter s along the search curve
                to produce a periodic function of  s as shown
                in Figure  6(b).    Note that the  sampled points
                shown here do  not include  the  extreme  values of
                c^  (points B and C').
     In  principle,  one  could  determine the  slopes  of the

solution surface  along  each  of the parameter axes  at a large
number of points  on the solution surface above  the domain of
uncertainty of the  parameters and compute suitable

                          52

-------
averages.  Such brute force methods, indeed, have been used
on occasion for systems involving only a few parameters and
small numbers of coupled differential equations.  The FAST
algorithm, on the other hand, provides a technique for
selecting much smaller numbers of sampling points.  The
method is most likely to be successful for those systems in
which the solution surface is smooth and preferably mono-
tonic, with minimal structure in the way of deep valleys,
sharp ridges, or peaks close to potholes.
     The method consists of four steps.  First, a set of
M integer frequencies o)£ is selected subject to certain
constraints to be discussed below.  Second, one frequency
is assigned (arbitrarily) to each of the parameters considered
to be uncertain, setting

         P£ = p° exp (u^Csin w^s ) }                   (3.35)
where s is a parameter of variation; p° is the "best value"
of p?; and u? is a continuous function of its argument
selected so that its extreme values result in values for
p. at the extremes of the range of uncertainty for p^.  Note that
p.£ is then periodic in the parameter s; as s  varies, the point
£ determined by  (3.35) moves about in the domain  of  uncer-
tainty of all the parameters.  If the set of  frequencies
co£ are mutually  prime, the point £ traces out (exactly  once)
a closed curve  (a Lissajous  figure) within the domain of
uncertainty as  s ranges  from 0 to 2ir.
      Third, the  values of p^ resulting from  setting
s =  2-nq/H*, q =  1, 2, 3,...., N*, are used  in  running  the air
quality simulation model from given initial conditions to
time  t.  The values of the solutions c . (_x,{ p (s ) } ,t)  are
stored after each simulation run. II* is proportional to
the  largest frequency to  .
                       A/

                          53

-------
w
O
     max
    jnin
     i
                    t/2
               TT

               S
                                      3TT/2
                                         2lT
    Figure 6(b).
The solution c.(-{p(s)}) is a periodic  function
of the search parameter s. The Fourier amplitudes
of this function at the frequencies assigned  to
any parameter p. provide an estimate of the
average slope of the solution surface  in the
direction of the p.-axis.The search curve does
not in general reach the extreme values of c.
        Fourth, the solutions c. for a  given  position  and time
   (x.»t) are now periodic functions of  s  itself  (Figure 6(t>).)  Fourier

   analysis of this function yields amplitudes  corresponding to

   each of the integer frequencies to^  :
                 2ir

        Ad>  = fc ^  ci({p(s)})sin  (iHs)  ds   .             (3.36)
          £      0

   Only sine components occur in the expansion  of c^(-p(s) )
   because of  the symmetries of Eq.  (3.35).   This integral can
   be approximated by a simple summation  over the accumulated

   soluti ons:
A
2
^ £  ci({p(sq)))sin
   q=i
                                      sq)
                                       (3.37)
   The Fourier amplitude  A    measures  the  contribution made by
   variations in  p^  to  the  total  variations  in c.,-  as all the
   parameters p oscillate back  and  forth  across the domain of
   uncertainty.
                              54

-------
     It is thus not surprising that the Fourier amplitudes
defined above can be shown to be proportional to a  weighted
average of the partial derivatives of c^ with respect to each
parameter:

             max    max     max        .
            Pi     Po      PM          3c.
            A      A       "       P(p) —L dp, dp,...  dpM
            t
            v)
            p-n   pjm    pnnn


         a < 7   >                                      I 3 38)
             •J0'                                   \ -J • 
-------
  mm
  max
  nan
                                                            '(P,)
Figure 7.   The functional forms used to generate the search path
            in the FAST algorithm are highly arbitrary, resulting
            in a great deal of ambiguity about the probability
            distributions P(p^).
                             56

-------
     The number N* of simulation runs  required is determined
by the highest frequency in the set {w^K  since each cycle
of a sinusoidal wave must be sampled at several points in
order to determine the amplitude of that Fourier component.
For economy,  we would like N* to be as  small  as possible; for
accuracy in evaluating the amplitudes,  we  would like N* to be
as large as possible.  If we make N* larger,  then we can also
increase tom,  . the largest frequency in the set, which has
          in a x
the effect of folding the search curve  back and forth in the
domain of uncertainty more times, increasing  the likelihood
of detecting  fine structures in the solution  surface and of
the search curve coming very close to  the  highest and lowest
points (the extremes) of the solution  surface.
     Various  algorithms have been developed for generating
sets of frequencies for use with FAST,   The frequencies should
be chosen to  minimize ambiguities in th.e determination of
Fourier amplitudes which arise from interferences between
frequencies or harmonics,  As a minimum, the  frequencies must
be linearly independent;
               M
               t   a^ f 0                      (3,39)

for  integer  coefficients  an-  subject to
                M
               £_    a.  <  Mj  +  1,                  (3.40)

where Mj  is  an  integer.   Such  a  set of  frequencies will  then
be  free of interferences  to  order Mj,
      Because  the  Fourier  amplitudes are computed  from  a  finite
number N*  of  equally  spaced  sampling points,  "aliasing"  can
also  confuse  the  picture.   Aliasing occurs whenever  a  sum of
two  or more  frequencies  (and/or  their  harmonics)  is  equal
to  an integer  multiple  of  the  number of sampling  points;
               M
               Y   a', to.  =  JN*                    (3.41)
                f=l   n   n

                           57

-------
where the coefficients a- and J are also integers.   To
minimize such aliasing, we require
       M
          ai  eoj  j  JN*                            (3.42)
wi th
       M
      L
             < M+1.                               (3.43)
The frequency setdo^} will then be free of aliasing to order
My.  For MT = 4, it has been found empirically that the  maxi-
mum frequency and  the number N* of sampling points  necessary
                     ? s
are proportional to M   . No analytic expression has yet been
obtained for N* in terms  of M and MJ-, such a formula would
have to be derived by the techniques of combinatorial algebra
which would have to provide first of all an explicit algorithm
for generating  the set of frequencies {w,}.
     The important features of the  FAST method are:
        •   It is a non-dedicated  method which  requires no
            alterations of the air quality  models.
        •   It provides relative measures of expected sensitivities.
            To date, only  first-order measures,  , have  been
                                                  I , K
            obtained in applications, although  higher order  informa-
            tion is potentially  available   should  there be a  need
            or interest in abstracting it.
        •   Applications  to  date   have emphasized  pure  kinetics  or
            pure non-reactive  hydrodynamics.
        t   Consideration  of  correlated  parameters  is in  the  early
            stages .
        •   Computer economy  of  individual  air  quality  simulation
            models  is  a major  factor  in  whether or  not  a  sufficient
            number  of  cases  can  be accumulated  as  adequate input
            to  the  FAST method.
                           58

-------
Variances obtained by summing all  specific frequency
components are generally inexact since only a finite
number of terms can ever be computed numerically.
The method can ultimately serve the interests of
both policy-makers and of scientific researchers.
Each application of the FAST method provides
sensitivities at a specific time point, in snapshot
fashion, rather than continuously  in time as do the
deterministic methods. (The same is true for other
sampling methods, as we saw for the Monte Carlo
method. )
In difficult problem regimes, the  results obtained
depend upon specific sampling distributions, as
is evident from considering Figure 7 and as
                              (19)
observed in practice by Penner;-  ' If the solution
surfaces are very irregular, different search patterns
yield different output results.
Given a sufficiently dense search  pattern, the method
can in principle resolve subtle system sensitivities
associated with complex wrinkles or pathological features
in the solution surface. (The extent to which this
possibility is ever realized in practice is problematical
due to considerations of computer economics as discussed
in Section 4.)
                 59

-------
                                         TABLE   1
                                         SUMMARY  TABLE

CIIAKACTtmSTICS  OF CAMDIDATt  STSTtHMIC  SCHSIlltllt  AMD  IIHCmAlll Tt ^HAHSIS  METHODS

METHOD VARIABLES

OCUHHiHlSllC KTHQOS «C( (x.t).
(A.) 1!ME-HO»0«£D Z( j, z"{ „








(B ) VARMTIOHAL 7, ,, Ac, ( first
'•J 'order)



• Curstant parameters
(direct method)
(1) pure Kinetics

(11) hydrodynamics
(reactive and
non- reactive)











• variable para-
•cters 7 functional
derivatives]
(1) pure kinetics


(11) pure hydro-
4yna«ics

SAtf-HHG KTHOOS
(A) Monte Carlo o.(x.t)
1 ~






PROGRAMHIHC FEATURES

Nun-Dedfcattd-
Rpsults are obtained as
continuous functions of
time for all variables.







Dedicated- Results ar*
obtained as continuous
functions of tine for
all variables. Sensiti-
vity vodel should be de-
coupled fron air quality
model.
(1) Highly automated

(11) only 1-D transient
sensitivity models
have be"r<4fls are r.uw
reasonably near at
hand with Highly
automated features
due to recent ad-
vances In PDC
•ethods.



(1) Seal-automated


(1i) Not-automated


Non- Dedicated.
Results ar« obtained at
single time points for
all variables




APPLICATIONS TO DATE

Models frcMi essentially all
scientific and engineering
disciplines—pure kinetics,
! f>, 2-0, and 3-0 transient
rodel*.






Only puie kinetics and 1-0
models. Ho Myher dimensional
sy.Uas.



(1) atmospheric chemistry.
chemical laser*, and com-
hustlon rnenfstry
(11) 1-D atmospheric dynamics.
further developments can
be readily made In 10,
whereas scwcwhat nor*
extensive developmental
•f f 01 ts art 3 Oil re-
quired For 2-D applica-
tions.









(1) diurnal photochemistry for
the simple Chapman
neclianisB.

(11) Test proble«s for a
dlffuslvt boundary layer.


1-D ste*4? state stratospheric
ozone Model





REQUIRED INPUT AND
EUCUTION fOH M PABA-
H£Tf US
Hlnlnuff of ( MM) itr
qualfty node! Integra-
tions-Includes both
nominal and perturbed
paianeter values. All
paidineters can be tested.










Che integration of th*
air quality rwxlel using
nominal parajtetcr values
plus M Integrations of
sensitivity equations.











The above Integrations
plus solutions of adjoint
sensitivity equation! are
are required, plus some
double Integrals *u$t be
computed. In addition to
being more costly, these
operations are difficult
to automate for arbitrary
problem scenarios.


On tbe order of (y) ^
nodel Integration*,
where Y Is the nunfcer of
saaf>Te points per para-
«eter anrt K. 1$ ch«
nukaber of sensitive para-
ncters. Generally Y'% 2
and K^ 10 for computer
•conol.y. All »tode1 paca-

COHMEHTS

This fs the favored wthod of mat
piactfcal tisers. No special pro-
nr*iw; are nrqulnd. The trlal-
and-error concept Is Intuttlvcl/
appealing and logically trans-
parent to » •jjotlty of users.
No statistical constdeiatlons art
required. All »od*l paramters.
whether constant or varying can be
tested. Results are a»*.iablc ta
best and worst case Identification.
Coming Into widespread use as at/to-
Mted proqiaw becune opplirahle
to arbitrary problem ictnarlos.
Requires no pre-screenfng of para-
neters for atmospheric kinetics.
No statistical ccms l.lei ationi arr
required, and 2. , and or, (f1n,t
order) are conceptually trans-
parent » Apptars to be mure ecorm-
nlcal than all other alternative
nethixJs (In both human effort >*4
computer costs). Results are
amenable to best and worst case
Identification.











Functional derivatives require
significant hands-on control by
the user, and are considerably
more costly to solve than cases
with constant parameters.
In practical applications It wovld
be Biost reasonable to approxlaut*
time variations by plecewlse coat-
slant variabilities to that a nra
efficient direct method can be «.-.*d
In place of functional derivatives.-

Results an* expectation values
whlcfl ate generally dependent u»«m
assuiwd paraaicter distributions.
Pre-screenlng Is necessary to
estimate In advance the nuaf>«r ef
sensitive parameters which In tun
govern the rate of convergence of
the method (istvjls). Results an
good to all orders and are concett-
                                                                           can be tested, but  unity  transparent.  R«uHs art mot
                                                                     applications to datt
                                                                     eftfriiaslze cht*«!cal rates.
                                                                                                      amenabU to bsst and worst ca»
                                                                                                      Identlftcatlon  uiilcss raw samp It-
                                                                                                      data are additionally (n-ocessed.
                                                                                                     'Even with prc-screenlng estlMtes,
                                                                                                      tlte otmfcer of satuples required U
                                                                                                      give convergence Is not fnown
                                                                                                      exactly  for each specific ?|(x.O;
                                                                                                      nonitorlng Is retjulrtd during
                                                                                                      evaluation of o '«,
partial varla

and  € lt  „ >
Non-Dedicated.  Results
art obtained at slngli
time points for all
vat tables.
                                      Hodali from several           On tlte order of 
-------
                         SECTION 4

        EVALUATION OF CANDIDATE SENSITIVITY METHODS
            FOR AIR QUALITY MODEL APPLICATIONS

     This  section  illustrates "what you get"  and "what
you pay" with  systematic sensitivity and uncertainty
analysis methods in their present states of development.
Evaluations of the candidate systematic sensitivity
and uncertainty analysis methods are given in the
discussion below in terms of our detailed assessments of
the performance characteristics of each method for:
(i)  chemical  models, (ii) 1-D transient models and (iii)
3-D air quality models.   The results are self-evident and
indicate a mixed picture--- one in which both reasonable
and unreasonable options are avaiable,  Accordingly, several
criteria are used for our evaluations.;
     •  Moti vat ions.  Fortunately, "what you get"  can be  very
        much a function of what you want or need in choosing
        systematic sensitivity and uncertainty analysis
        methods. Optional methods are available for those
        with poli cy-maki ng motives and  for those in atmos-
        pheric research.  These respective motivations will
        largely determine the selection of candidate analysis
        methods for air quality simulation models.
     •  Cost.   This can be the make-or-break  factor,  in-
        volving not only computer execution costs  but also
        human programming costs.  Human programming costs
        include the development of new  sensitivity programs
        (for dedicated  methods) as well as the essential
        tasks of preparing input data  and  writing  preprocessor
        and graphics routines.  For large, complex regional
        air quality models the human costs alone can prove
        to be prodigious- depending  upon  specific applications,
        as can computer execution expenses.  This  is true

                             61

-------
both for dedicated  sensitivity programs,  per se,
and for the multitudes  of air quality model  runs  which
may be needed as input  to non-dedicated sensitivity
methods.  For smaller,  purely photochemical  kinetics
box models or for purely hydrodynamics models,
computer costs are  more reasonable as are human
programming costs when  automated compilation
features are used.
Useful ness. We particularly focus on criteria which
underlie whether or not potential users may  sooner
or later be motivated to use candidate systematic
sensitivity and uncertainty analysis methods.  In
addition to labor and computer economy, it is essential
that a method be conceptual1y transparent and that its
results be readily interpretable.  (This factor alone has
accounted for a majority of practical users  continuing
to stand by the time-honored method.  The overwhelming
appeal of the time-honored method derives from several
factors;  it is highly intuitive in nature,  it requires
no special preparations to operate  initially,  and no
special operations are needed to interpret output
results.)  An additional criterion of utility is
the potency of  candidate analysis methods.  Here,
the major promise of some systematic sensitivity
methods is their ability to identify more effectively
and to evaluate quantitatively subtle sensitivities
in complex solution surfaces.  Potent methods also
have  the capacity to anticipate  unexpected events
[new  discoveries or other surprises.)
Whether or  not  potency  of a method  should be  associated
with  the  availability  of  higher-order  sensitivity  measures
is  a  subject  of continuing  debate.   On the  one hand,
policy  makers  desire reliable  variances  (uncertainty
measures)  for  the  results  of  air quality  models.   The
                    62

-------
        most  reliable  means  of  obtaining  such  information
        is  by  sampling  results  from  repetitive runs  of  the
        air quality  model  (basically,  the time-honored  or
        Monte  Carlo  method).   It  is  clearly  out of the  range
        of  practicality to construct variances by  first
        evaluating  and  then  summing  terms involving
        Z-  -,  Z-  .. , etc,,from  either the  deterministic  or  the
         1 »J    1 » J K
        FAST  methods.   On  the  other  hand, users who  require
        sensitivity  measures are  generally quite satisfied
        with  first-order sensitivity coefficients  (regardless
        of  the methodology which  is  used  to  generate the
        coefficients)  because  first-order coefficients  contain
        all of the information which most users want and need
                                                             *
        on  a value-for-yalue basis  \n practical applications,
     With some reflection, all of these criteria are
recognized  to be branches of the same tree,  the tree of user
acceptance.  This should not be surprising because these
issues of user acceptance have perpetually governed the rate
of development and the pattern of use of systematic sensitivity
and uncertainty analysis methods. Tables 2 and 3 provide a
summary of our evaluations of the candidate analysis methods.
Detailed discussions of these evaluations follows directly below,
BASIC CASES
     Three  basic examples are used for purposes of illustrating
evaluations in this section.  In  all cases,  computer cost
*  This view has been expressed repeatedly by the overwhelming
   majority of scientific practitioners; it represents, in fact,
   the hard reality of the situation.  This does not say that
   higher-order sensitivities are insignificant—we have clearly
   shown the contrary in previous sections--but rather that, on
   a value-for-value basis, first-order sensitivity coefficients
   provide a major part of the information which practical users
   seek from sensitivity analysis.

                           63

-------
estimates assume that a Control  Data Corporation (CDC) 7600
computer is used for sensitivity and uncertainty computations
at a cost of $1200 per execution hour ($20/minute).   This is
obviously  an arbitrary basis, but it can be converted approxi-
mately to  the appropriate, equivalent basis for other computer
systems.   The main  points of our evaluations should be quite
evident within  this  selected context despite the great vari-
ability of  computer  execution times  for the various models
and computers which  are  used by different investigators.

A.  Chemical Models  (purely kinetics)
     This  base  case  assumes that:
     •  Thermal  and  photochemical rates are constant.
         (Time-varying  rates are far  more costly  to  integrate
         using existing ODE system solvers)
     •   External  chemical sources are constant.
     f  The system  has N  chemical species and  M  reaction
         rate parameters.
     •   Numerical  integration of the ODE system  of  kinetic
         (species)  equations has costs which are  proportional
         to Na .  It  is  reasonable to  assume  that  a^2-3,  which
         is representative of implicit solution methods for
         stiff ODE  systems, such as  variants of the  Gear  method "'
         Generally,  a system of  20 species would  require
         approximately  15 seconds of  CDC  7600 execution time
         at a cost  of about $5 per kinetics  run;  a  system of
         50 species  requires about 4  minutes  CDC 7600  execution
         time at an  approxinate  cost  of  $80  per kinetics  run.
         Kinetics execution times are not very  sensitive  to  the
         number  of reaction rate parameters, M.
      •   Automatic kinetics programs  are  available  which
         allow  users to select,  compile,  and solve  arbitrary
         reaction sets  (with  arbitrary  rates),  requiring
         virtually no dedicated  programming.  "
                            64

-------
B.   1-D Transient Atmospheric Models
     One-dimensional  transient models have been applied               |
primarily to reactive plumes and to global stratospheric              •
ozone modeling.   In addition, 1-D models provide a convenient         i
foundation for subsequent considerations of 2-D and 3-D               |
regional air quality simulation models.   This base case                j
assumes that:                                                         i
     •  thermal  and photochemical rates are constant over
        intervals of one hour or longer.  (Instantaneous
        variations greatly increase execution costs).
                                                                      i
     t  External  chemical sources are constant  over
        intervals of one hour or longer.
     •  The system has N chemical species, M parameters,
        and L spatial zones.
     •  The cost  of numerical integrations  depends  upon  the
        choice of solution methods.  (For example, simultaneous
        space-time solutions using  conventional matrix
        methods  are far more costly than space-and-time
        split solutions using Gauss elimination methods.)
        Current  models tend to require on the order of
        several  minutes of CDC 7600 execution time per
        simulation when simultaneous  space-time solution
        methods  are used.
     •  Automatic compilations of arbitrary chemical
        mechanisms, velocities, and diffusion profiles
        are possible, but are not yet in widespread use.
     •  N = 20,  L = 50, M = 150.   (We assume there are
        50 chemical rates, 50 velocity values, and 50
        diffusion coefficient values  at the respective
        grid points,  resulting in M = 150.)  Initial and
        boundary  conditions will  be considered in later
        examples.
                             65

-------
C_.   3-D Regional  Air Quality Models
     This base case assumes  that:
     •  Thermal and photochemical  rates  are constant over
        intervals of one hour or longer
     •  The spatial grid is  25 x 25  x 5,  for a total of
        3125 elements.
     •  N = 10
     t  The system has 25 chemical rates, of which 5 are
        photodi ssociation.
     t  3 velocity components (measured  at 3 hour intervals)
     •  Diffusion  is characterized by horizontal  and  vertical
        components KH and KV,
     e  The model  treats 5 emission factors (NO  , CO, 3
                                               J\
        hydrocarbons), measured hourly.
     •  Initial  values are available for all species, the
        temperature, and the  relative humidity   (R.H.).
     •  Boundary  values are  available at all surfaces.
     0  While  numerical integration costs  are highly
        dependent  upon choices  of solution methods, current
        models of  this type  (which split space and  time
        integrations) require approximately 30 minutes of
        CDC 7600  execution  time for a 24 hour simulation.

 CHEMICAL  MODEL EVALUATION

 A.   The  Time-Honored Method
      (i)   Sensitivities
      What  can  be  simpler  than  merely  altering  the  value  of
      any parameter  p-, while holding  the  remaining  parameters
                     J
      at  their  nominal  values,  and  directly  noting  the  resulting
      variations  in  all solutions,  6c-(t)  for  itl,..N,   con-
      tinuously in time?   This  is  the  essence  of  the time-
                           66

-------
honored method.   The performance characteristics of
this method are:
•  Statistical  properties of parameter distributions
   are not required.
•  Values of 6c-(t) are accurate to all orders.
•  Values of 6c-(t) can be obtained with equal- facility
   for time-varying rate coefficients as for constant
    r^tes.  (The cost of diurnal  kinetics runs can exceed
    constant parameter runs by factors  of 5 to  50.)
•  A minimum of (M + 1)  kinetics runs are  required--
   one run for  the nominal  parameter set and one run
   for each rate  which is varied.   (Assume  all  M are
   to be varied.)   In practice,  several trial  values
   may be required in order to locate the  sensitive
   range of some  of the  parameters,  which  frequently
   leads to a  cost of ^ 2M kinetics runs. For  20  species
   and 50 rate  parameters,  the cost of 2M  runs would
   be approximately $5  x 2  x 50  =  $500.  For 50 species
   and 100 rate parameters, the  corresponding  cost
   would be $80 x  2 x 100 -$16,000 (for testing all
   100 parameters).
•  Potentially  new discoveries or  unknown  rates can
   be tested by including hypothetical reactions in the
   basic mechanism and then determining values of their
   rates which  would be  required in order  for  the
   hypothetical  reaction to become significant in the
   overall mechanism.
•  Whenever new nominal  rate values are used or new
   reactions are  included in a mechanism,  the  sensitivity
   analysis must  be repeated in  toto.
                       67

-------
   Sensitivity coefficients can be obtained as
   Z.  .  =  Urn   A-c-i ,
    1,J    Apj+0 Apj '
                      A2C.
   l.(2}   - lim        	1—-   ,  etc.
    1 >Jk   Apj ,Apk->0   Ap,Apk

   Recall, however, that it is very poor economy to
   evaluate sensitivity coefficients as  a means of
   obtaining 6c.j  (see  Eq.  2.3) to high-order accuracy.
   It  is far more effective to get Ac-, directly —by simply
   introducting any desired perturbation Ap- and using
                                           J
   the time-honored method.
•  Expectations of sensitivity coefficients,
   ,  can be obtained by accumulating a huge
     1 , K
   number of outputs from kinetics runs in which all
   permutations of simultaneous rate perturbations
   have  been introduced.   This requires on the order
     r  ]M
   of [ Yj  kinetics  runs  where y is the number of
   trial perturbations to be applied to each para-
   meter.  If only two trial values are used for each
   of 50 parameters, |2|     kinetics runs would be
   required, which is clearly out of the range of
   p r a c t i c a 1 i ty .

(i i)  Uncertainties
0  Variances a- can be evaluated from data accumulated
   in the same fashion as was mentioned immediately above
   In this case Stolarski et aP15' have found that the
                                        " i M
   number of samples is much less than  JY| ''•  Tn1's wil1
   be discussed further below under Monte Carlo Methods.
                      68

-------
     •  Best case and worst case estimates for 6c1- can be
evaluated by directly  applying selected combinations of para-
meter perturbations.   The choices of parameter perturbations
are arbitrary,  and the process of finding best and worst
case parameter  combinations is not unique; exhaustive trials
appear to be required.

B.  Variational (Direct) Methods
     These methods are most appropriate for evaluating sen-
sitivity coefficients Z-  *. The performance characteristics
                       i »J
for first-order sensitivity computations are:
     •  The sensitivity equations (3.3) can be automatically
        compiled and solved for arbitrary chemical mechanisms
        The solutions can be obtained either simultaneously
        with the chemical kinetics equations or decoupled
        from them.  Decoupling must be carefully performed
        and thoroughly validated against  the simultaneous
        case;  this is  the subject of current research.

     •  The costs for  decoupled sensitivities are about
                                                   3
        the same as a  kinetics run, which goes as N .  This
                       o
        amounts to (MN ) for a system of M parameters
        plus one kinetics run at nominal parameter values
                                                       •3
        which gives a  total cost proportional to (M+1)N  .
        For 20  species and 50 parameters, this implies a
        cost of approximately $250; for 50 species and 100
        parameters, the approximate cost is  $8000.  The
        costs of solving the sensitivities and kinetics
                                               o
        simultaneously is proportional to M(2N)  for all
        its parameters, or about 8 times as expensive as
        the corresponding decoupled case.  The approximate
        costs of the simultaneous case are $2000 for 20
                          69

-------
         species and  50  parameters,  and  $64,000 for 50 species
         and  100 parameters.   Clearly  the redundant  kinetics
         solutions which  characterize  the  simultaneous case
         should be avoided.
     •   Costs of evaluating  systems with  time-varying
         coefficients  are  inordinately  greater, both  in
         machine execution  time  and  in  dedicated programming
         of  functional  derivatives.
     •   Whenever new  nominal  values or  new  reaction  sets
         are  used, the  sensitivity analysis  must be reperformed.
     •   Expectations  of sensitivity coefficients  
                                                    i »j
         are  not practical  to  evaluate.   Estimates of 6c-  and
        Oj  based on  first-order sensitivity coefficients  have
        on "y local validity  (see Section  1.)
 C.  Green  s Function Method
    This r-ethod is  similar in many  respects to the variational
 (direct) nr^thods  just discussed, but  has not been tested  or
applied as  extensively as yet.   Key performance  characteris-
tics of th is method  are as follows;
    •   The  homogeneous differential  equations  (3.30) for the
       Gre -en's  function are easier  to solve (in  terms  of
       cor-,putation time) than are the sensitivity equations
       (3.  3) of the  direct method,  and can  be  automatically
       corn-foiled and  solved for arbitrary   chemical  mechanisms.
    •   The   numerical solution of (3,30) must be  reinitialized
       rep-- eatedly,  adding some expense to the  computation.
       Add  itional  expense is incurred in computing  the  sen-
       sit- ivitiesZi  k(t) from the Green's function  by
      qua_ drature  (trapezoidal rule or' equivalent).   These  added
      cos- ts, however, should be smaller than  the savings
      aqh- ieved by  decoupling the sensitivity  equations  from
      the   kinetics  equations.
                           70

-------
     •   Whether  or  not  the  decoupling  itself  can  be  done
        without  losing  accuracy  remains  to  be  proven  for
        reasonably  complicated  kinetics  systems.
     •   If  desired,  higher  order sensitivity  coefficients  can
        be  obtained  at  far  less  cost  than by  the  direct
        method.
     •   Additional  programming  efforts  totalling
        perhaps  one  person-year  should  allow  the  develop-
        ment  of  a  user-ready  package  which  could  resolve
        these questions.

D.   The Monte Carlo Method
     This is  the extension of the time-honored method
which computes variances of model output data.  Performance
characteristics  are as follows:
     t  This  method is conceptually transparent;  it is
        non-dedicated; and it applies to one time point
        at a  time.
     •  ["Y]   kinetics runs are required as  input, where
        Y is  the number of trial perturbations applied to
        each  parameter.  For j=2, the following dollar
        costs would be incurred for (constant parameter)
        kinetics runs alone:

        20 Species
        50 parameters  -  $ 5.6 x 10^ ^ 1nf1n1te
        100  parameters -  $ 6.3 x 10  ,

        50 Species                   -,
        50 parameters  -  $ 9 x 10" ;  ^ 1nf1n1te
        100  parameters -  $ 1 x 10   |

        For  all  practical purposes, these costs  are  infinite,
        suggesting that more selective parameter sampling
                          71

-------
must be used.  One alternative would be the FAST
sampling scheme.   So far as we can determine, this
possibility has not yet been explored in a practical
example.  We will see  in (E) directly below that
even the FAST sampling would be inordinately
expensive in kinetics integration for our basic
                           (IS)
examples.  Stolarski  et  alv  ' estimate  in advance
the likely order-of-magnitude cost of variance
evaluations.  They apply the time-honored method,
using trial perturbations of +0 and +_ 2 a for all
rates,  in order to make a preliminary identification
of the  most sensitive parameters in the system.
(Recall that this is a local evaluation, varying
one rate at a time about nominal parameter values,
which is good to arbitrary order and is continuous
in time.)  In the event that the kinetics mechanism
is sensitive, say, to only 10 rates out of the
original set of 50 or 100 rates, then only (2)   or
1024 kinetics integrations are required as input
for  Y=2.  In this event the cost of the input for
computing variances would be $5120 for 20-species
systems and  $81,920 for 50-species systems.  This
approach is  now a naturally-evolved hybrid of sen-
sitivity and uncertainty analysis.  This practical
alternative  will prove useful for all of the systematic
methods . Whereas in the Monte Carlo method,  prescreening
only serves  to estimate execution costs, in  the  FAST
method, prescreening  followed by rejection of insensitive
parameters may result  in considerable savings of input
samples without compromising  the basic results.  This
point  should be further pursued andeither  validated
or  invalidated by  FAST users.
In  cases where prescreening  is  used for parameter
rejection,the  hybrid  approach exacts  a cost  elsewhere,
                  72

-------
       namely, some of the promise which systematic methods
       hold for a more rigorous search of model  solution sur-
       faces is compromised by computer economics,
    t  The entire analysis must be redone if any nominal
       values change.   In some cases,  changes in the magnitude
       of parameter uncertainties (probability distributions)
       may also require that the entire analysis be redone.
    •  Sensitivities tested with perturbations of +a do
       not sample the  "corners" or extreme regions
       of the solution surfaces.  Best case and worst
       case examples are more likely to escape notice
       in these circumstances.
    0  Stolarski  et al found that output variances were
       essentially independent of the  insensitive parameters.
       But the rate of convergence to  reliable variances
       was dependent upon the degree of sensitivity of
       particular variables in various time (and space)
       domains.  Hence the user of the Monte Carlo method can
       not know a priori the number of input runs which are
       required to obtain convergence  of variance values,
E.   Fourier Amplitude  Sensitivity Test

     The  performance characteristics for first-order
sensitivity coefficients are:
     •  The FAST is a  non-dedicated method which can be
        applied to arbitrary model outputs at one time
        poi nt of the model.

     •  The  number  of  kinetics  integrations  which are
       required   as input  to  the  FAST  is  approximately
       (M)2-5.   Therefore,  approximately  17,700  kinetics
       runs  are  required  for  50  parameters;  and  100,000
                          73

-------
   Nineties runs are required for 100 parameters.
   The following dollar costs would be incurred for
   (constant parameter) kinetics runs alone:
   20 Species
   50 parameters  - $  88,500
   100 parameters - $ 500,000
   50 Species
   50  parameters - $1.42 x 10
   100 parameters - $8. x 10
t  It is clearly necessary to minimize the number of
   parameters under consideration as well  as to minimize
   the cost of reliably solving kinetics systems.

 •  A conundrum arises.  The  promise of  systematic
   methods  is to search the  model solution  surfaces
   more  thoroughly  in  order  to detect subtle as  well
   as obvious reaction coupling  influences.  Increasingly
   thorough  searching, in turn,  suggests  that  more
   and  more  complete  reaction  mechanisms  should  be
   put  to  the test.   But  the computer economies  argue
   strongly  in the  opposite  direction,  thereby  thwarting
   some  of  the promise of systematic methods which
   happen  to  be  expensive.
 t  Variances  can be obtained by  two means:
      (i)   The output  of  the numerous kinetics  runs which
   are  required  by  the FAST  can  be  processed directly.
   It is known that the parameter space is  sampled  non-
   uniformly, depending upon the specific  choice of
   FAST  distribution  functions.  The  effect of sampling
   choices  upon  such  output  variances has  not  yet been
   assessed,  but is likely to  be significant for sen-*
   sitive  parameters.
                        74

-------
  (ii)   The  FAST obtains  the  sensitivity measures
  from  partial  variances  a • (p -) =a . (w .  t)/a. .   The
   »J                           '  J    '   J »     i
 terms  c-(u>.  t)  are  defined  in terms  of  the  Fourier
        1  j j
 amplitudes  at  frequency  w.  by
                          J
 In  practice  only  a  finite  number  of  to.  and  associated
                                      J
 harmonics  are  actually  used,  so that the  resulting
 variance values may or  may not be  reliable.

  Unaveraged sensitivity measures  Z-   • (analogous
                                   •  > J
  to measures obtained by the deterministic methods)
  can be obtained in principle by  using distributions
  in the FAST of zero width (zero  uncertainty) for
  all Pi^i* but this would be very inefficient relative
  the other sensitivity methods  for kinetics systems.
  In principle, second-order and perhaps higher-order
  sensitivity information is available from the FAST.
  In practice, this capability does not yet appear to
  have been implemented for computations by general users
  As discussed above, this additional information
  will not provide  a more effective means of obtaining
  improved variances vis a vis other alternatives;
  and the application of higher-order, averaged
  sensitivity information appears  to be a very low
  priority need of  most practical  users, due to
  problems of conceptual transparency and ease of
  application.  (It is our impression that this is
  the case for higher-order sensitivity information
  from all existing methodsj
                    75

-------
     •  The sensitivity measures  depend in  general
                                   1 5 J
upon the choices of distribution functions,  as was discussed
                              fl.8)
above and also found by Penned  ; in  applications.  The
sampling methods thus tend to be best-suited to problems in
which probability distributions  are well-known for key parameters

1-D TIME DEPENDENT MODEL EVALUATION
     From the evaluation of chemical  models  just above,
it is clear that computer economics will  dominate  our
evaluations of sensitivity methods for  1-D and higher-
dimensional transient air quality models.   The cost of
integrating multi-dimensional transient models can be
highly variable and is particularly dependent upon the
numerical methods and programming efficiency in any specific
air quality model.   To provide a basis  for subsequent
evaluations we will first discuss briefly  a  set of well-
known numerical solution options which  are used in multi-
dimensional transient models and then indicate approximate
scale relationships of the costs of these  solution options
to the costs of purely chemical  models  which were  previously
stated.
     Transient 1-D chemical-transport models are generally
solved simultaneously in (.x,t) by employing  the method of
lines to discretize spatial gradient operators.  An implicit
stiff ODE integration method is  then applied to solve the
overall  system of size  NL x NL| .  When no special numerical
advantages are exploited, the cost of solving such a  1-D
model is proportional to rNL|a,  where a has  a value between
2 and 3.  When central differences are  applied to gradient
operators, resulting  in an overall matrix having a block-
tridiagonal structure; Gauss elimination solutions of the
matrix can be used which lead to reduced solution costs  on
the order of LNa.
                         76

-------
     JBy similar analysis, simultaneous  solutions  in  (x..y_»t)
in 2-D transient models  (e.g.,  LIRAQ) have solution  costs
ranging from L3Na to  [L2Nj a.   (For  the  remainder  of  this
discussion we can assume a ^ 2-3, but the structural  aspects
are clearer if we Scale  costs  in  terms  of a.)   The use  of
ADI (alternating direction implicit solution)  methods for
2-D transient models reduces the  numerical solution  cost
to the range of 2L2Na to   2LJLN]a.  In  3-D air quality
models (e.g., the Systems Applications,  Inc., regional  model), ADI
is absolutely essential.  Solution  costs for 3-D  air quality
                                                f\
models would then scale  to span the range of 3L Na to
3L  LNJa (for Gauss elimination versus  conventional  matrix
solution options).!
     For estimation purposes,  we  will assume that 1-D models
are solved simultaneously in (_x,t)  at a cost of 5 minutes
of CDC 7600 time, or $100 per  simulation.  Again, this
estimate is representative  of existing transient  1-D models
with which we are familiar; the main thrust of our con-
clusions will not be greatly affected by this   particular
choice of cost estimate  vis a  vis execution costs of other
such 1-D transient atmospheric  models actually in operation
today.
A.  The Time-Honored Method
     Assume that, in addition  to  the single model solution
which  is required using  nominal parameter values, only
one additional run will  be required for each parameter  in
order to evaluate  6cn-(x,t).  (This  might  be the  case for  a
                   I "~~~
very experienced practitioner.)  A minimum number of  para-
meters would be M = 150, which  is composed of:

     •  50 chemical rates which are constant in time and
        have the same value in  every zone.  Assume that the
        same rate perturbation applies  simultaneously to
        every zone.                 '
                           77

-------
     t  50 diffusion  coefficients  which  are  uncorrelated
        and take  specific  values  at  each  zone.   Assume
        that a  single fractional  perturbation  at each
        zone is appropriate  over  the entire  time history.
     •  50 wind velocities which  are uncorrelated and
        take specific values  at  each zone.  Assume that
        a single  fractional  perturbation  at  each zone  is
        appropriate over the  entire  time  history.
     In practice,  many other  parametric  variations may
also be considered, including
     t  pollutant emission sources at each zone as a
        function  of time;
     t  boundary  values  of species concentrations or
        species fluxes as  a  function of  time;
     •  initial conditions at every  zone; and
     •  photodissociation  rate variations from zone-to-zone
        due to  variable  cloud cover  (even for  constant rate
        coefficients).
     On this minimum basis,  the  cost of  1-D air quality
model runs alone  would be  (151 x $100) or $15,100 of com-
puter time.  (The reader can  readily make similar estimates
which are germane to his own  computing circumstances and
experiences.)  Major improvements in computer economy could
be realized from possible  advances in four areas.  First,
improved numerical techniques such as matrix solutions by
Gaussian elimination would probably  bring the  computational
cost under $1000 for 150 model runs.  Second,  intuitive pre-
screening for obviously  insignificant sensitivities can reduce
the number of parameters to be varied; the utility of this
option depends  strongly  upon the craft of the user.  Third,
correlated variables can sometimes be identified  and varied
                           78

-------
simultaneously rather than individually.   Fourth, optimized
programming, using automatic compilation  of arbitrary problem
scenarios, machine coding, parallel  processing, and other
specialized program optimization routines, can result in
                                                  (15)
substantial savings.   For example,  Stolarski et aP ' found
it necessary to write a completely  new stream-lined production
version of their original 1-D atmospheric model and further,
to reduce it to a steady-state model  rather than a transient
model.   However, this latter option  would probably not be
appropriate for multi-day air quality simulations which are
intrinsically transient in nature.
B.  Variational Methods
     Using the same basic example as  above, several operational
possibilities can be dismissed immediately.  For example,
simultaneous solution of the sensitivity  and 1-D model equations
is out of the question.  (Recall the discussion for Chemical
Models.)  Similarly, the sensitivity equations would be solved
for one parameter at a time rather  than solving the larger
matrix for, say, all M parameters simultaneously.  Further,
automatic compilation for arbitrary  problem scenarios is
essential for both the sensitivity  programs and 1-D air
quality models.  (Can you imagine the difficulty of writing
a new dedicated sensitivity program which corresponds per-
fectly to each air quality model and/or to each specific
application which is of interest?)
     Following these suggestions and assuming that solutions
of the sensitivity equations would  cost the same to obtain
as solutions of the 1-D air quality  model, one would be solving
a system of size [ML x Nil (M+l) times.   A practical approach
to calculating sensitivities for this problem (with constant .
chemical rates) might be to pursue  the following lines.'
                          79

-------
     •  For a given set of nominal  v° and KR  at each grid
        point, solve the sensitivity equations MR times for
        the reaction rates of interest.   This gives Z- -L.
        at all grid points continuously in time.
     •  For a given set of nominal  chemical rates, solve
        the sensitivity equations Mv and M,,  times for the
                                   -      =r>
        velocities and diffusion coefficients at selected
        grid points of interest.  (v^ and KD may be constant
        over, say, hourly intervals; this would pose no
        difficulty to the direct methods.)  This procedure gives

        Zi,J Ivel.  and Zi,J 'oiff.  at a11 System 9rid
        points, continuously in time.
     o  For a given set of nominal  values for chemical  rates,
        winds, and diffusion coefficients, the sensitivities
        Z- .  can  also be determined at  all grid points, con-
         's J
        tinuously in time, for selected  initial, boundary,  and
        source conditions.
    So far as we  can determine, no  sensitivity  analysis of
this  level of generality  has yet been performed with  direct
methods.   (Koda et al^  have considered  much  more  limited
examples,  such as  non-reactive  atmospheric  diffusion,  which
could be  solved analytically.)  In  the  event  that  M=150 in
the general  case  (due  to  any combination  of  values  for Mn,
M  , M.,  ,  and  other parameters), the  computer  cost  would be
 —•    — D
identical  ($15,100)  to  the  minimum  case  discussed  for  the
time-honored  method  immediately above.  Notice that the
possibility  of using intuitive  or  other judgements to eliminate
obviously insensitive parameters from the systematic analysis
is  tacitly left open, recognizing:
      (i)   the element of  risk in such eliminations, and
      (ii)  the fact that some of the  promise of  systematic
methods is compromised  to computational economy by  such
eliminations.

                           80

-------
      The functional derivative approaches for either time-
 varying rates or non-discretized spatial  variables  are  severely
 limited by the following factors:

      0  The methods are too highly dedicated, at present,
         for multi-dimensional application.
      •  1-D model execution costs increase dramatically
         for diurnally varying rates and continuously varying
         emissions sources.
      0  We estimate execution costs for solving the air
         quality and the adjoint sensitivity equations to
         be on the order of solving an [ML x Nil  system
         of ODE's for the air quality equations plus  an
         [N2L2 x N2L2] system of ODE's for the adjoint
         equations.  Thus the execution cost would scale
                          r  o o^
         approximately as [N Lja, with a relatively  large
         proportionality factor applying in the case of
         variable rates, sources, etc.  In addition, N
         double integrals at,say, L values of x^ would have
         to be evaluated at each time point for each chemical
         rate.  Likewise, N  double integrals for each
         pair (_x' , xj and each time point would have to  be
         evaluated for sensitivities to winds  and diffusion
         coefficients.   We conclude that,  vis  a vis  other
         methods, this approach is far too costly and difficult
         to apply in air quality modeling for the foreseeable
         future.

C.   The Monte Carlo Method
     For 150 parameters, (2)    runs of the 1-D model is,
practically, an infinite input cost.  To date this method has
been used primarily to obtain steady state atmospheric
model  variances associated with uncertain chemical rates.
Even in the extremely unlikely event that only 10 parameters
were found (by some other means) to be sensitive, 1024 runs of
a 1-D transient model would  be on the order of $100,000 in
computer execution cost.  This is the reason that
               f 15)
Stolarski et a 1      emphasize the need to take every
                             81

-------
reasonable step to reduce the execution cost of the atmos-
pheric model which generates the basic data for the Monte
Carlo method of evaluating variances.   In the event that
20 sensitive parameters are present in an air quality
model, over a million  input samples would be required to
evaluate variances, which is out of the question if air
quality model execution costs greatly  exceed 1/100 cents
per run.
D.  The FAST Method
     The input cost of testing 150 parameters would be
% (150)2'5 % 275,500 runs of the 1-D air quality model,
costing on the order of $27.5 million  for computer execution.
The non-dedicated  features of this method remain very
attractive,  but the number of parameter trials must be
markedly reduced  either by pre-screening eliminations or
by considering only segments of the overall problem (such
as testing  the chemical mechanisms independent of the
fluid  dynamics or  vice versa).

3-D AIR QUALITY MODEL  EVALUATION
     The general  picture is abundantly clear by now.   If
we consider only a 24-hour simulation  with the 3-D model at
a cost of approximately $600 per run,  the computer costs
relative to the 1-D examples present a grim picture,  indeed.
To carry through this discussion,  Tables 2 and 3 have
presented  what we  havetermed an "Oversimplified Case"  and a
"Realistic Case" for the number of parameters which must be
considered in our  basic 3-D air quality model.
 The  "Oversimplified  Case"  is  simplified  beyond
scientific usefulness.   The "Realistic Case" is scientifically
acceptable but is  well  out of the range of practicality.
The results  in Tables  2  and  3  are  self-explanatory.
                           82

-------
                                                     TABLE 2
                                                 SUMMARY  TABLE
                                 BASE CASES FOR ILLUSTRATION  OF EXECUTION COSTS
                             OF CANDIDATE SYSTEMATIC  SENSITIVITY ANALYSIS METHODS
             BASE CASE
                                     DESCRIPTION
                                                              ASSUMED EXECUTION
                                                              COSTS  (on basis of
                                                              "typical" CDC 7600)
                                                                                     COMMENTS
             PURE CHEMICAL
             KINETICS MODELS
   (A.) 20 species, SO
   parameters

   (B.)  50 species, 100
   parameters
$5/klnetlcs
simulation

$80/klnetlcs
simulation
The paramatera
are assumed to be
constant.   Diurnal
and other time
dependent effects
would be treated by
using piecewise
constant parameters
over intervals of an
hour or longer.
             1-D TRANSIENT MODELS
                                     20 species, 50 grid
                                     points.(Hence there
                                     are 50 diffusion
                                     coefficient values
                                     and 50 wind values on
                                     the basis of one value
                                     per grid point and
                                     uncorrelated y_(x) and
                                     KD(x).)
                                                              $100/1-0 simulation
                                                   Parameters  are
                                                   assumed  at  least
                                                   piecewise constant
                                                   (as above).
               3-D REGIONAL AIR
               QUALITY MODELS
10 species,
3125 grid points  (based
on zoning which Is
dimensioned as 25 x  25
x 5),

25 chemical rates,

3 velocity components,

"Wir. and "Srert.!

5 emission factors (NO  ,
CO, three reactive
hydrocarbons, measured
hourly),

Initial values for
species, temperature,
and relative humidity.
Boundary values at all
surfaces.
                                                              $600/3-D simulation
A.  Over-simplified Case.  137 parameters, assuming
that:

•  v_ has identical values at all grid points.

*  ^H* ^v nave identical values at all grid points.
•  No cloud corrections; thus J's do not vary from one
   grid point to another.
•  k's do not vary from one grid point to another
   jCconstant T, R.H.  at all grid points)

•  Initial conditions for 10 species, T, and R.H. are
   identical at all grid points.

•  Boundary conditions of v_, KU»KV.  and 10 species are
   identical over each surface. (15 components x 6
   surfaces - 90 values)
                        B.   Realistic Case. *\» 234,200 parameters, assuming that:


                        •  _v la variable at each grid point.

                        •  KH and Ky are variable at each grid point.

                        •  5 Jfs, which are variable at each grid point.

                        •  20 k's, which are variable at each grid point.

                        •  Initial conditions are variable at each grid point.

                        •  Boundary conditions are variable at each grid point.
                                                         83

-------
                              TABLE 3
                             SUMMARY TABLE
ILLUSTRATIVE  COMPUTER EXECUTION COSTS FOR CANDIDATE SYSTEMATIC
SENSITIVITY ANALYSIS METHODS (REFER TO BASE CASES IN TABLE  3.1(a))
METHOD

TIME-HONORED
(6Ci(t))






DIRECT VARIATIONAL
(Z, ,(t))
i , J









FAST
(^ ,-(!')>)
1 »J


MONTE CARLO
(O.j(t'))





PURE CHEMICAL
KINETICS MODELS

(A. ) $500 (Assumes
two trials
per para-
meter)

(B. ) $16,000




(A.) $250 (Assumes
sensiti-
(B.) $8000 argyd^S'
coupled
from kine-
tics egs,)

(A.) $2000 (As^H™6?
(B ) $64, 000*1 "-V"*
eqs. are
coupled)
(A.) $88,500 (Assumes
, no pre-
(B.) $8.10b screen-
ing)


(A.) $5.6 x 1015
(B.) $l.xl032 $82-200(assume
B One
(B.) $1.4xlOB trial /
para-
meter)
(13 yrs CDC 7600

-------
                         SECTION 5

                     TODAY'S APPROACH

     The central  objective in this section is to suggest
the means by which the EPA, today, could most effectively
apply systematic  sensitivity analysis methods to arbitrary
air quality models, which we will assume to be transient
2-D or 3-D regional or reactive plume models.  Our suggested
choices of method will be guided to a significant extent
by the conclusions which have evolved from our evaluations in
theprevioussections.

CONCLUSIONS
     A.  On balance, the full promise of systematic sen-
sitivity analysis methods cannot yet be realized in applications
to entire 2-D and 3-D transient air quality models.  However, by
segmenting the models or by prescreening for sensitive para-
meters, very useful sensitivity information can be gained with-
out si gni fi cant alterations being required in the air quality
models.

     B.  A combination of time-honored and systematic sen-
sitivity analysis methods can be effectively used on individual
air quality model segments; for example just the chemistry can
be analyzed in arbitrary zones with the fluid dynamics remaining
decoupled, and conversely the chemistry can be frozen while
analyzing the fluid dynamics, the boundary conditions, the
initial conditions, sources, etc.  (Without question pre-.
screening and segmenting compromise   some of the promise of
systematic methods in exchange for computational economy;
this compromise is necessary until further advances occur in
systematic sensitivity analysis and/or in the efficiency
of execution of air quality models.)
                           85

-------
     C.   Reliable variances  appear to be out of practical
range for multidimensional  transient models, with a multitude
of sensitive parameters.
     D.   Dedicated sensitivity methods must be programmed
so that arbitrary problem scenarios can be automatically
compiled in a hands-off manner; otherwise, non-dedicated
methods are a far more practical  alternative for analysis
of complex air quality models.
     E.   For better or for worse, practical users favor con-
ceptually transparent sensitivity methods over those which
require more subtle interpretation without providing com-
mensurately more useable information.  (For example, higher-
order sensitivity coefficients hold important topological
information about complex solution surfaces; but the volume
of information which is needed to accurately map complex
surfaces grows quickly to unmanageable and conceptually
obscure proportions.)
     F.  Numerical efficiency of air quality models and
sensitivity methods assumes ever-increasing importance when
extensive sensitivity studies on the models are desired.
(Most air quality models would require considerable re-
programming to reach their maximum attainable efficiencies
for  production purposes.)
     6.   Depending on specific circumstances, field or
laboratory measurements maybe a preferred alternative to
sensitivity analyses.  At one extreme are the chemical models
where sensitivity analysis is very cost-effective for
indicating the most critical reactions which need to be
measured in the laboratory in order to markedly improve a
chemical mechanism.  At the other extreme are 3-D regional
models where the costs for even a si ngle sensitivity analysis
                          86

-------
may easily exceed $100,000 for computer expense, alone.
But we have seen that, in addition, many such analyses
must be done to account for alternative parameter choices
which are furthermore subject to many permutations.  In
this case, comprehensive uncertainty analysis of the 3-D
model is unreasonable and should be supplanted by more
meaningful experimental work.  Between these extremes,
however, more subtle tradeoffs must be very carefully
considered on an individual case basis.

TODAY'S PRACTICAL APPLICATION OF SYSTEMATIC SENSITIVITY ANALYSIS

      Variances  of  air quality model  outputs can be economically
 and reliably  obtained for only very  special circumstances
 (such as  models with less than about 10  sensitive  variables
 and for highly  specialized models  which  require only seconds  or
 fractions of  seconds of CDC 7600 execution time
 (or equivalent) per simulation).   This  conclusion is
 equally true  for air quality model  segments: For example,
 variances from  a purely chemical  model  (or submodel),  in-
 dependent of  other segments of an  air quality  model, can
 be reliably obtained only for cases  in  which about 10   or
 fewer  chemical rates are sensitive  parameters.   We thus
 believe that  policy-making applications  will be best served
 by shifting the emphasis  away from studying variances  of
 limited scope and  validity and,  instead,  to increased
 understanding of the physical and  chemical cause-and-effect
 mechanisms with the aid of hybrid methods  of sensi ti vi ty analysis,
      Today's  best  approach to sensitivity analysis of complex
 air quality models would  first decompose  the subject model
 into a chemical component and a hydrodynamics component,
 noting several  additional factors  which apply to the model
 segmentation  process:
                           87

-------
     (i)  Segmenting  is  most  effective when it is applied
     at the points  of weakest mutual  coupling in the air
     quality model.   The chemical-hydrodynamics decoupling
     is an obvious  first choice,  based on present experience;
     improved  decoupling points  should always be sought.
     (ii) Carefully  considered application of model segmenting
     requires--and  returns—greater understanding of the chemistry
     and physics  in  a given model.
     (iii) If incorrect assumptions  are made in model seg-
     menting,  quite  often the result  is that more--not  less--
     is learned  about both  the air  quality model and the
     simulated air  quality problem.

Analysis of Sensitivities in  the Chemical Component
     The chemical mechanism of any  large air quality model
can be readily  simulated by stand-alone automated chemical
kinetics solvers  such as those described in reference 17,
Arbitrary conditions  (e.g.,initial  conditions,  species
fluxes, dilution  factors, and pollutant sources and sinks)
can be  abstracted from any spatial  zone and time domain of
the subject air  quality  model and provided as input for
solution by the  more  efficient purely  chemical  kinetics
solvers.  This  can  be done repeatedly  for representative
conditions and  circumstances  in  the original air quality
      *
model.
     Photochemical  sensitivities can  be determined most
economically if  rate  parameters  are at least piecewise
constant in time  (over one-hour intervals, for example)
rather than instantaneously varying on a diurnal basis.
* This process  of abstracting such conditions from a more general air
  quality model to a more specialized (photochemical) model is known as
  "linking".  Linking has repeatedly been demonstrated to be  effective
  in validating complex computational models in numerous scientific
  disciplines.

-------
(Instantaneous  diurnal  variations  exact such an  extreme
cost in kinetics integrations that all  sensitivity methods
can become unduly expensive.)  Recognizing that  the
application of  a kinetics sensitivity method is  very  much a
matter of  user  preference, we will nevertheless  indicate
as a guiding suggestion that our present choice  would be
to  use a  direct variational method on  the chemical com-
ponent for several  reasons:

     9  Economy of effort and computer  cost.  If decoupled
        from the kinetics solutions,  a  suitably  automated
        direct  sensitivity method  is  more economical  than
        the other alternatives in  terms of  human effort and
        computer cost,  particularly for new applications
        to large chemical mechanisms.  Data handling  is
        minimized  as are arbitrary operational  decisions
        such as trial and error of selection of  rates,  rate
        distributions,  etc.
     •   Conceptual  transparency.   The measures 6c-(_x,t)
        of the  time-honored  method and  Z-  .(t) of  the  direct
                                        ' »J
        method  are  both obtained continuously  in  time  and
        are independent of statistical  assumptions, unlike
        the measure   of the  FAST.   The time-honored
                      • > J
        variable  6c-(x,t)  has an advantage  of  being valid
        to all  orders.  From experience  with all  of the
        alternative sensitivity  methods  (except  for the
                           89

-------
Green's function method which is not yet imple-
mented for general  use) we have usually found that
the automated direct method gives the basi c sen-
sitivity information which is needed in a more
organized and efficient, hands-off manner than the
alternative methods.  (Again, individual user pre-
ferences may very well  differ from ours.)
Potency.  Direct variational  methods come closest
to fulfilling the full  promise of systematic sen-
sitivity analysis methods in  chemical applications.
No pre-screening (usually done by trial and error)
is required in atmospheric applications of direct
methods, whereas pre-screening is advisable for the
FAST when more than about 20  parameters are to be
tested.  Pre-screening should always be minimized
to the greatest possible extent because any over-
sights in this stage will compromise all subsequent
evaluations.  The measures Z-  . and 6c-(x,t) of the
                            1 > J        i —
direct and time-honored methods can also be used for
best and worst-case identification, whereas 
                                              ' »J
from the FAST suppresses such  information by the
averaging process.  By reverse reasoning, the FAST
is the most effective method for providing averaged
sensitivity coefficients in those special instances
where  the averages  have meaning.  It may also occur
to some that the FAST  can effectively  compute Z-  •
                                               • > J »
as well as  ,  by  simply assigning  zero uncer-
              1 > J             f h
tainties to all except  the j~Ln parameter.  But we
have seen that, in  the  absence  of new  advances,  this
apparent option of  the  FAST  is  grossly  inefficient
relative to the direct  variational  and  time-honored
alternatives,  particularly for  moderate  to  large-
size chemical  mechanisms.
                   90

-------
    •  Ease of Application.  A majority of practical users
       prefer the trial and error approach and tend to resist
       more formal sensitivity analysis methods which require
       their own numerical programs.  The great intuitive appeal
       and immediacy of the time-honored approach ensure that
       it will continue as a favored sensitivity method.  On
       the other hand, direct methods have now been developed
       and automated to a  sufficient degree^' ^' * ' for all
       types of chemical model  use that widespread application
       should be near-at-hand.  Koda et al^' are similarly
       reducing the FAST method to semi-automatic general
       applicability in atmospheric chemistry problems.

     The sensitivity testing of the purely chemical component
 is  likely to be most meaningful if the systematic analysis (by
 the direct method) of rate and initial condition uncertainties
 is  augmented by independent trial-and-error assessments of
 factors such as instantaneously varying photodissociation rates,
 more detailed characterizations of time dependent sources and
 sinks, and any other factors which were either by-passed or
 treated approximately by the direct method.

 Analysis of Sensitivities in the Hydrodynamics Component
     Whereas  automated chemical  kinetics  solvers  can  be used
 efficiently  to assess  essentially  unlimited chemical  scenarios
without running  the  full  air quality model,  the same  is not
 true for  the  hydrodynamics  component.   The problem is  that,
 even for  non-reacting  fluids,  independent hydrodynamics solvers
 are not feasible  for higher than one-dimensional  transient
 models  with  arbitrary  zoning,  sources, sinks, and wind and
 diffusivity  resolutions.   Similarly, direct sensitivity
 methods have  not  yet been effectively automated for arbi-
 trary scenario evaluations  in  higher than one-dimensional,
 transient hydrodynamics.   It is therefore necessary for the
                          91

-------
original  air quality model,  or some trivial  derivative of it,
to be used to generate necessary output data for sensitivity
analyses  of hydrodynamics factors.   For example, the chemistry
can be turned off by assigning zero rates to all chemical
reactions in order to study  the purely hydrodynamic (meteoro-
logical)  characteristics of  the air quality model.
     With extensive basic understanding of the physical
problem,  the time-honored method would then be used almost
exclusively to identify the  key variables and their basic
sensitivities to space and time dependent meteorological
parameters, sources, sinks,  initial conditions, and boundary
conditions.  The time-honored method would serve also as  a
means of pre-screening insensitive variables before applying
the FAST to small groups of  carefully selected hydrodynamics
parameters.  Correlated parameters (i.e. those which are re-
presented by common functional relationships at all spatial
zones) may also be treated effectively by the FAST in particular
circumstances.  Clearly, basic understanding of the physical
system,  coupled with sound intuition, is  the most important
element  for incisive sampling and successful analysis of
hydrodynamic sensitivities;  for it is simply impossible to
test every potentially sensitive parameter at every spatial
grid point with any of the presently developed sensitivity
methods .

NEEDED BREAKTHROUGHS  IN SYSTEMATIC SENSITIVITY ANALYSIS METHODS

      Breakthroughs  which  would provide  truly  major advances
 in sensitivity  analysis  applications  fall  into  two major
 categories;  first,  there  are  general  advances  which  would
 benefit  al1  specific  sensitivity  and  uncertainty  analysis
 methods, and second,  there  are specialized  advances  which
 will  benefit,  exclusively,  a  particular sensitivity  method.

                          92

-------
General  Advances
     •  Model  execution speed.   We have seen that several
sensitivity methods  would become more attractive if the cost
of air quality simulations were on the order of 1/100 cents
per run.  However,  this possibility appears to be out of
reach for even perfectly optimized 2-D and 3-D transient
air quality models  with envisioned computers of the.next
generation  or two.
     •  Advances in partial differential equations.
Significant advances in PDE numerical analysis are likely
over the next 5-10  years which  may dramatically improve air
quality modeling.   First, numerical integration of non-
linear systems of PDE's is.in  the midst of major advances.
Finite element methods with adaptive gridding features •'
have the capacity to use far fewer grid points, located at
optimal positions.   These features tend to eliminate numerical
diffusion and the overshooting  errors which have always
plagued the finite  difference methods which are used in
current-generation  models.  These new approaches are, additionally,
very amenable  to automatic compilation of arbitrary problem
scenarios in higher dimensions, in a manner which  is very
much akin to the automated kinetics  (ODE) solvers which were
mentioned  in  previous  sections.   These  general  advances un-
doubtedly will be many  years in coining  to  fruition,  unless
special  initiatives are  undertaken.
      •   Correlated  variables allow significant  reductions  in
execution costs for all  systematic sensitivity  analysis methods.
The  identification  of  correlations in winds and diffusivities ,
for  example, will most  likely  come from very  fundamental  considera-
tions of fluid dynamics.  Such studies are far  more basic  and
rigorous than most  air  quality models, per se,  and will thus
require many more years  of highly detailed theoretical and
experimental analysis.

                           93

-------
Specialized Advances
     •  The FAST method would be markedly improved by more
efficient parameter sampling which relates to the selection
of incommensurate Fourier frequencies.
     •  The FAST method would run markedly faster if
correlated parameters can be identified and represented by
a common parametric variation, which in turn reduces the
required number of  independent sample points.
     0  The direct  variational methods are markedly improved
when the sensitivity computations are decoupled from the
kinetics or hydrodynamics computations.  A very modest amount
of research (about  l person-year) will now accomplish the
development of reliable decoupling methods.
     •  The direct  variational methods will advance markedly
when 2-D hydrodynamics models can be automatically compiled
for arbitrary problem scenarios  and solved by generalized
hands-off  PDE solvers (such as is presently the case with
automatic  kinetics  solvers which use the Gear ODE solution
method^1*  ^8').  Additionally, the sensitivity equations
must be amenable to the same measures.  At Science Applications,
Inc.,  in Pleasanton, California, in-house efforts are, in fact,
remarkably  near (within 3  person-years) of accomplishing this
advance using very  potent finite element methods.   (We are
aware  of only one other such effort,  in a different field,.
which  is near a  comparable point of advancement,  so it is
difficult  to  give a more  universal status estimate than the
one above  which  is  based  on our  own work.)
     • Finally  the Green's function  method may  ultimately
offer  some  advantage over alternative  sensitivity methods
for purely  chemical systems.   It is still  too early to tell.
 (We estimate  that  1-2 person-years may yet be required to
satisfactorily   implement this method  in  a suitable' automatic
format for arbitrary mechanism evaluations.)
                              94

-------
                    REFERENCES


 1.   Reynolds,  S.D.,  P.M  Roth  and  J.H.  Seinfeld,  "Mathematical
     Modeling of  Photochemical Air  Pollution  -I.  Formulation
     of  the Model.", Atmos.  Environ.,  7,  1033-1061.  (1973).

 2.   Mac Cracken,  M.C.,  D.J. Wuebbles,  J.J.  Walton, W.H.  Duewer,
     and K.E. Grant,  "The  Livermore Regional  Air  Quality  Model:
     I  and  II".  J.A.M.. 17,  No.  3,  254 (March  1978).

 3.   Sklarew, R.C.,  M.A.  Joncich,  K.T.  Iran,  "An  Operational  Air
     Quality  Modeling  System for  Photochemical  Smog in San  Diego
     Air Basin",  Fourth  Symposium  on Turbulence,Pi ffusion,and Ai r
     Pollution,  301,  (AMS  Jan  15-18, 1979).

 4.   Hecht,  T.A.,  J.H.  Seinfeld,  and M.C.  Dodge,"Further
     Development  of  a  Generalized  Kinetic  Mechanism for Photo-
     chemical Smog", Environ.  Sci.  Techno!.,  Vol.  8, No.  4,
     pp. 327-339  (197TT

 5.   Whitten, G.I. and  H.  Hogo,  "Mathematical  Modeling of
     Simulated  Photochemical Smog", EPA Report  EPA-600/3-77-011
     (Jan.  1977).

 6.   Gelinas, R.J. and  Skewes-Cox,  P.O., "Tropospheric Photo-
     chemical Mechanisms",  J.  Phys. Chem. .  81,  2468 (1977).

 7.   Dickinson,  R.P.  and  Gelinas,  R.J., "Sensitivity Analysis
     of Ordinary  Differential  Equation  Systems",  Journal  of
     Computational Physics, Vol.  21, No. 2,  123-143, (June  1976).

 8.   Hindmarsh,  A.C.  and  G.D.  Byrne,"EPISODE:   An Experimental
     Package  for  the  Integration  of Systems  of  Ordinary Differential
     Equation Systems",, UCID-30112 ,  Lawrence  Livermore Laboratory
     (May,  1975).

 9.   Atherton,  R.W.,  R.B.  Schainker, and E.R.  Ducot, "On  the
     Statistical  Sensitivity Analysis.of Models for Chemical
     Kinetics",  A.J_._C  h_.J.  J_. ,  21.,  441  (1975).

10.   Gelinas, R.J.,  "Diurnal Kinetic Modeling", Proceedings  of
     the International  Conference  on Structure, Composition  and
     General  Circulation  of the  Upper and  Lower Atmosphere  and
     Possible Anthropogenic Perturbations.  Vol. II, Jan.  14-25,
     1974,  IAMAP/IAPSO,  Melbourne,  Australia,  (1974).

11.   Koda,  M.,  AH.  Dogru,  and  J.H. Seinfeld,  "Sensitivity
     Analysis of  Partial  Differential Equations with Application
     to Reaction  and  Diffusion  Processes",  preprint (1978).
                            95

-------
12.   Walter,  W.,  Differential  and  Integral  Inequalities,
     (Springer-Verlag,  New York,  1970).

13.   Porter, W.A., Int. J. Control. _5,  393 (1967).

14.   Hwang, J.T., E.P.  Dougherty,  S. Rabitz, and H. Rabitz,
     "The Green's Function Method of Sensitivity Analysis in
     Chemical Kinetics", preprint (1978).

15.   Stolarski,  R.S.,  D.M. Butler, and R.D. Rundel , "Uncertainty
     Propagation  in a  Stratospheric Model, 2.  Monte Carlo
     Analysis of  Imprecisions due to Reaction Rates", J6R,
     in press (1978).

16.   Cukier, R.I., C.M. Fortuin, K.K. Schuler, A.G. Petschek,
     and J.H. Schaibly, J. Chem. Phys.  59, 3873,  (.1973).

17.   Dickinson,  R.P. and Gelinas, R.J.,  "SETKIN",  in L. Lapidus
     and Schiesser (eds.), Numerical Methods for  Differential
     Systems, (Academic Press, 1976) , pp  167-180.

18.  Penner, R.C., private communication,  also see A.A. Boni
     and R.C. Penner,  Combust. Sci. Technol. 15,  99, (1976).

19.  Doss, S., private communication (_1978),
                             96

-------
                                  TECHNICAL REPORT DATA
                           (Please read Instructions on the reverse before completing)
  REPORT NO.
   EPA-600/4-79-035
                                                          3. RECIPIENT'S ACCESSIOf*NO.
 4 TITLE AND SUBTITLE
  SYSTEMATIC SENSITIVITY ANALYSIS OF AIR QUALITY
  SIMULATION MODELS
5. REPORT DATE
     May  1979
6. PERFORMING ORGANIZATION CODE
 7 AUTHOR(S)
  R.  J.  Gelinas and J.  P. Vajk
                                                          8. PERFORMING ORGANIZATION REPORT NO.
9 PERFORMING ORGANIZATION NAME AND ADDRESS
  Science Applications, Inc.
  80  Mission  Drive
  Pleasanton, CA  94566
10. PROGRAM ELEMENT NO.
  1AA603A AB-042 (FY-79)
11. CONTRACT/GRANT NO.

  68-02-2942
 12. SPONSORING AGENCY NAME AND ADDRESS
                                                          13. TYPE OF REPORT AND PERIOD COVERED
  Environmental  Sciences Research Laboratories-RTP,  NC
  Office of Research and Development
  U.S.  Environmental Protection Agency
  Research Triangle Park, N.C.  27711	
  Final   1/78 - 12/78
14. SPONSORING AGENCY CODE
  EPA/600/09
 15. SUPPLEMENTARY NOTES
 16. ABSTRACT
       Systematic sensitivity and uncertainty analysis  methods are reviewed and
  assessed for applications to air quality simulation models.   The candidate methods
  in terms of their basic variables, mathematical  foundations, user, motivations and
  preferences, computer implementation properties,  and  costs  (including both human and
  computer resources) are discussed.  Both deterministic  and  sampling methods have been
  evaluated with illustrative examples of "What you pay"  and  "What you get" with present
  generation systematic sensitivity analysis methods and  air  quality models.  Determi-
  nistic methods include the time- and user-honored methods of arbitrary parameter
  adjustments by trial and error, variational methods,  and a  newly formulated Green's
  function method (for kinetics systems only).  Sampling  methods include Monte Carlo
  sampling of the outputs of air quality models to  compute variances and a Fourier
  analysis method of sampling model outputs to compute  expectation values of sensiti-
  vity coefficients.
       Computational economics, inclusive of both  programming effort and computer exe-
  cution costs, were found to be dominant governing factors for the effective applica-
  tion of systematic sensitivity and uncertainty analysis methods to arbitrary air
  quality problem scenarios; several reasonable options and several unreasonable option'
  emerge from the present evaluations.  Recommendations are made outlining how EPA
  should, today, most effectively apply available  multi-parameter systematic sensitivity
j17 analysis metnoas TO arDitrar^^-^^n^^t^u^^d^^udj ,ty ^.nuiau.u,, ,,,uucia.
'a. DESCRIPTORS jb. IDENTIFIERS/OPEN ENDED TERMS
* Air pollution
* Mathematical models
* Analyzing
* Probability theory
;13 DISTRIBUTION STATEMENT
i
. RELEASE TO PUBLIC

19. SECURITY CLASS (This Report)
UNCLASSIFIED
20. SECURITY CLASS (This page)
UNCLASSIFIED
c. COS AT I Field/Group
13 B
12 A
14 B
21. NO. OF PAGES
105
22. PRICE
EPA Foi-m 2220-1 (9-73)
                                            97

-------