Office of Superfund Remediation and
Technology Innovation
Sediment Assessment and Monitoring Sheet (SAMS) #2
Understanding the Use of Models in Predicting
the Effectiveness of Proposed Remedial Actions
at Superfund Sediment Sites
Phenanthrenc Flux Q Site A
dav) Q Sit« 3
Surface
Dtgradation
:
OSWER Directive 9200.1 -96FS
November 2009
-------
"^\ Understanding the Use of Models in Predicting
the Effectiveness of Proposed Remedial Actions
at Superfund Sediment Sites
SESSMENT AND MONITORING SHEET #2
Background and
Purpose
This is the second fact sheet in the Sediment
Assessment and Monitoring Sheet (SAMS) series
prepared by the Office of Superfund Remediation
and Technology Innovation (OSRTI).
This product is a primer for those not experienced
in the development and use of models at sediment
sites. It explains the typical objectives of modeling,
how models are built, how they are used to predict
the effectiveness of remedies, and how the
uncertainty in model predictions can be addressed.
The document is not intended to provide site-
specific direction on the application or data
requirements of specific models.
This document does not supersede the guidance
on modeling provided in section 2.9 of the 2005
Contaminated Sediment Remediation Guidance for
Hazardous Waste Sites. That document provides
guidance on determining whether mathematical
modeling is needed and what level of modeling is
most appropriate for a site, and discusses the need
to verify, calibrate, validate, and peer-review
models.
This document does not impose legally-binding
requirements on EPA, states, or the regulated
community, but suggests modeling approaches that
may be used at particular sites, as appropriate,
given site-specific circumstances.
This factsheet has been prepared by the U.S.
Environmental Protection Agency (EPA) Office of
Superfund Remediation and Technology
Innovation. Drafting and revisions were provided by
environmental modelers at LimnoTech under
subcontract with TetraTech EMI (Prime Contract
Number EP-W-07-078).
INTRODUCTION
Remedy evaluation at Superfund sites typically includes predictions
of risk reduction for each potential remedial alternative. As
discussed in the 2005 Contaminated Sediment Remediation
Guidance for Hazardous Waste Sites (USEPA, 2005), "Models are
tools that are used at many sediment sites when characterizing site
conditions, assessing risks, and/or evaluating remedial alternatives.
A complex computer model (e.g. multidimensional numerical
model) may not be needed if there is sufficient weight of evidence
distinguishing the best remedial option based on an adequate
understanding of site conditions; however, this is not often the case.
At some sites, significant uncertainties exist about site
characterization data and the processes that contribute to relative
effectiveness of available remedial alternatives. Models can help fill
in gaps in knowledge and allow investigation of relationships and
processes at a site that are not fully understood. For this reason,
simple or complex modeling can play a role at most sediment sites."
"Whether and when to use a model, and what models to use, are
site-specific decisions and modeling experts should be consulted."
http://www.epa.gov/superfund/health/conmedia/sediment/
guidance.htm
This fact sheet is an introductory primer for managers of
contaminated sediment sites who are seeking to better understand
the purpose and appropriate use of environmental models. It
explains how sediment models are built, calibrated, and used to
make predictions of remedial outcomes, how to decide how much
complexity to include, and also how to interpret predictions in the
light of uncertainties and other limitations of data and modeling.
Specifics regarding the development of data sets to support
modeling, selection of environmental models, or site-specific
application are not discussed in this document. The following
sections of this Fact Sheet address these questions about models and
their uses:
What is Modeling?
What'sin a Model?
What's Needed for a Model?
How Certain are Model Predictions?
How Do We Predict Remedial Outcomes?
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
1
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
important Principles to Consider in Developing and Using Mod
at Sediment Sites (USEPA, 2005)
1 Consider site complexity before deciding whether and how to apply a mathematical
model. Site complexity and controversy, available resources, project schedule, and the level
of uncertainty in model predictions that is acceptable, are generally the critical factors in
determining the applicability and complexity of a mathematical model. Potential remedy costs
and magnitude of risk are generally less important, but they can significantly affect the level of
uncertainty that is acceptable.
2. Develop and refine a conceptual site model that identifies the key areas of uncertainty
where modeling information may be needed. When evaluating if a model is needed and in
deciding which models might be appropriate, a conceptual site model should be developed
that identifies the key exposure pathways, the key sediment and water-body characteristics,
and the major sources of uncertainty that may affect the effectiveness of potential remedial
alternatives (e.g. capping, dredging, and/or monitored natural recovery (MNR)).
3. Determine what model output data are needed to facilitate decision making. As part of
problem formulation, the project manager should consider the following: 1) what site-specific
information is needed to make the most appropriate remedy decision (e. g. degree of risk
reduction that can be achieved, correlation between sediment cleanup levels and protective
fish tissue levels, time to achieve risk reduction levels, degree of short-term risk); 2) what
model(s) are capable of generating this information; and 3) how the model results can be
used to help make these decisions. Site-specific data collection should concentrate on input
parameters that will have the most influence on model outcomes.
4. Understand and explain model uncertainty. The model assumptions, limitations, and the
results of the sensitivity and uncertainty analyses should be clearly presented to decision
makers and should be clearly explained in decision documents such as proposed plans and
records of decision (RODs).
5. Conduct a complete modeling study. If an intermediate or advanced level model is used in
decision making, the following components should be included in every modeling effort:
Model verification (or peer-review if a new model is used)
Model calibration
Model validation
. Consider modeling results in conjunction with empirical data to inform site decision
making. Mathematical models are useful tools that, in conjunction with site environmental
measurements, can be used to characterize current site conditions, predict future conditions
and risks, and evaluate the effectiveness of remedial alternatives in reducing risk. Modeling
results should generally not be relied upon exclusively as the basis for cleanup decisions.
7. Learn from modeling efforts. If post-remedy monitoring data demonstrate that the remedy is
not performing as expected (e.g., fish tissue levels are much higher than predicted), consider
sharing these data with the modeling team to allow them to perform a post-remedy validation
of the model. This could provide a basis for model enhancements that would improve future
model performance at other sites. If needed, this information could also be used to re-
estimate the time frame when remedial action objectives (RAOs) are expected to be met at
the site.
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
WHAT IS MODELING?
Models and Their Uses
Starting with a Conceptual Site Model
The term "model" describes a broad range of
tools that can be used to integrate and analyze
available data. Models provide a framework for
understanding site behavior and predicting the
effects of actions taken at a site. A model can be
as simple as a statistical regression, and as
complicated as a process-based mathematical
description of the physics, chemistry, and
biology of a complex sediment site. Common to
all models, though, is the need for a conceptual
understanding of site behavior, i.e., a conceptual
site model (CSM), which is a representation of
the environmental system and processes
determining transport of contaminants from
sources to receptors.
At most Superfund contaminated sediment sites,
the environmental system includes the food
chain as a risk pathway for ecological receptors,
humans, or both. Aquatic organisms can be
exposed to contaminants in pore water,
overlying surface water, sediments, and through
their diet. Exposure to contaminants in
sediments occurs primarily in the top layer (an
"active layer" of sediment which can vary
widely but is often on the order of 10 cm in
thickness). Contaminants in deeper sediments
can also serve as a source of contamination
through their upward transport in pore water or
erosion and reworking of the sediments by
storms or other disruptive events that
reintroduce them into the active zone.
Contaminated sediment can be directly toxic to
biota and can also serve as an entry point for
food chain exposures to predators, anglers, and
hunters.
To effectively manage and reduce risks due to
sediment contamination, it is important to
understand the processes that brought about
those risks. This means understanding past and
ongoing releases, the transport of chemicals in
the environment, any changes in the form of
those chemicals over time, and the pathways of
exposure and risk to human and ecological
receptors. This set of relationships linking
releases to risk is included in the CSM. A CSM
is formulated for every Superfund site, and site
investigation supports the development, testing,
and refinement of the CSM. In turn, the CSM
can be used to identify gaps in the
understanding of a site.
A CSM identifies the processes that lead to
contamination and elevated risk and therefore
need to be considered in remedial planning.
These processes can be quantitatively
understood and incorporated into a
mathematical model, i.e., a set of quantitative
relationships relating model inputs (e.g., initial
conditions, boundary conditions, contaminant
inputs, hydrometeorology, or watershed solids
loads) to exposures and risks. While a
mathematical model can improve the
description of contaminant pathways and
underlying processes, its primary purpose is to
predict specific reductions in exposure and risk
from potential remedial actions. It is often
recommended that the CSM be developed into a
mathematical model where sites are large and
complex; where it is important to compare the
effectiveness of remedial alternatives over long
periods of time; and when a model offers an
opportunity to support decision making.
Quantifying a CSM: Mathematical Models
Mathematical models include analytical models,
regression models, and process-based numerical
models. Analytical models include universal
equations that provide a good fit to data without
need for calibration using site data. Analytical
models have limited applicability to
contaminated sediment sites because of the
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
complexity of processes and heterogeneity of
conditions. Regression models provide a best
statistical fit between independent and
dependent variables. They can be useful in
establishing relationships between variables
from past data, such as the relationship between
flow and suspended contaminants, or between
sediment and fish tissue concentrations, but are
limited in their power to forecast effects of
remedial actions that change the character of the
site, due to their reliance on past data. For these
reasons, process-based numerical models can be
the most useful mathematical models for
contaminated sediments.
When there is a need to describe or forecast site
behavior that cannot be captured with an
analytical or regression model, a process-based
numerical model is often applied. This is often
the case in contaminated sediment systems.
Sediment contaminants, especially organics
with low water solubilities, tend to be strongly
associated with solid particles, especially fine
particles and particles that have a high organic
carbon content. Those particles are often the
products of watershed erosion, ongoing erosion
of a river bed or banks, or naturally occurring
organic materials. As flow rates vary through
the year, those particles can be eroded from the
sediment bed and moved downstream, and they
can also be buried by other sediment. The
amount of contaminant that is adsorbed to
solids, dissolved, transported to the atmosphere,
or transformed into other chemicals can be
affected by a variety of physical processes,
including hydrodynamics, the nature of the
solids, temperature, pH, availability of oxygen,
and biological activity. A numerical model can
be used to describe those processes. Numerical
models that describe physical processes can be
combined with historical data on contaminant
concentrations to simulate past, present, and
future exposure concentrations at specific site
locations. A wide range of numerical models
exist that are capable of simulating physical
properties and processes such as water surface
elevations, velocities and shear stresses, solids
concentrations and fluxes, contaminant
concentrations and fluxes, and many other
variables that may vary across large areas or
long periods of time (Figure 1).
SS/C 22.374 i
555.574 i
Bedload 4,593 ,
1 f t
1 1 t
III * * 1
Contaminant/solids
mass balance
estimates
Contaminant
distribution maps
Contaminant trend
estimates (water,
sediment, biota)
Figure 1: Example Output from Numerical
Models Used to Simulate Various Physical,
Chemical, and Biological Processes
Adding Complexity as You Go: Tiers of Model
Development
As described above, the term "model" describes
a broad range of tools with varying levels of
complexity. Models can be classified into levels
of increasing complexity, or "tiers", which
provide increasingly detailed representations of
the physical, chemical, and biological processes
at work in a sediment system (Figure 2). The
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
first tier includes simple empirical and statistical
models useful for detecting statistically
significant trends in contaminant exposure and
exploring and testing for correlations among
environmental variables (e.g., river discharge,
temperature, water column contaminant
concentrations, etc.). However, simple statistical
models are inherently limited in their ability to
predict future conditions. Such models typically
are "fits" to available historical and
contemporary data, and as such are
unconstrained by the physics of the system
being simulated.
The second tier builds on the first by using
observations about trends and correlations,
combined with an understanding of basic
processes, to further develop the conceptual
model of the system. For example, it may be
possible to identify differences in the way
contaminants in carp and smallmouth bass trend
over time, and link those differences to known
differences in feeding habits, thus linking
differences to the degree to which each species
is affected by contaminant trends in the
sediment bed. Such observations superimpose
knowledge of the physical behavior of
contaminated sediment systems on empirical
observations of the site captured in a Tier 1
model.
Tier 3 modeling involves organizing the
knowledge of the masses of water, solids, and
contaminants in different system compartments
into a quantitative framework that measures
fluxes into and out of these compartments and
associated rates of accumulation in each
compartment. By quantitatively tracking mass
moving through the system, mass-balance
modeling helps answer questions that are critical
to evaluation of long-term trends, such as: What
is the rate of accumulation of solids in the
sediment bed? or, What is the rate of suspended
solids and contaminant export downstream?
The complexity of the site, the scope of
decisions to be made, or the specific
management questions asked may require a
more detailed modeling evaluation, often to
provide an improved understanding of some
critical piece of the system. A Tier 4 model is an
extension of the Tier 3 mass-balance models
that adds more detail or process modeling in
important areas. This might include detailed,
fine-scale and multidimensional
hydrodynamics; a more mechanistic description
of sediment transport and sediment bed
handling; or a more mechanistic description of
contaminant fate and transport processes. It
might include the addition of other supporting
modeling evaluations such as simulation of
wind-wave dynamics, wind-induced currents, or
extreme event modeling.
The decision to move from simpler to more
complex models requires careful consideration
of the need for and value of the added
complexity. Increased model complexity can
reduce decision-making uncertainty when used
appropriately and when well supported by site
data. If used poorly, added complexity can be
misleading and can even increase uncertainty. A
critical task for any modeling team, supported
by experienced modeling consultants and other
individuals not vested in the model
development, is to carefully select a level of
model complexity that provides real benefit and
is appropriate to the resources available and the
decisions being made.
Using Models
People who use and develop models often think
of them primarily as prognostic tools - a way to
predict the future conditions at a site. But when
developed along with a project, starting simple
and adding complexity as needed, models can
support many other aspects of a sediment site
investigation. Models from each of the different
tiers of modeling described in Figure 2 can:
Support directed data gathering during a
remedial investigation
Perform hypothesis testing and refine the
CSM
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
5
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Act as prognostic tools for predicting future
behavior of the system
Support evaluation and selection of
proposed remedies
Support remedy design
Help understand post-remedy monitoring data
Model Tier
Physical Understanding/
Strength of Forecast
No understanding of physical
behavior;forecasting based
on trends only
Initial understanding of key
process rates and
coefficients; forecasting
assumes those processes
Process rates and
coefficients further informed
by mass balance; forecasting
Integrated, mechanistic
hydrodynamic, sediment
transport, contaminant fate
and transport models
forecasting constrained by
detailed physics of
hydrodynamics, sediment
transport, chemical fate
Figure 2: Tiers of Modeling Complexity
These activities can span the entire timeline of
site investigation and remedy implementation.
Models, especially simple models, can be used
at the outset of a project to support planning for
data collection, in the middle of a project to
allow testing of important elements of the
system behavior, and at the end of a remedial
investigation to help choose between different
remedial alternatives. And when a model is
developed early, built with stakeholder input
and peer review, and employed throughout the
project, stakeholders can develop confidence in
the value and utility of the model. This
consensus will support its use when the project
comes to the critical point of remedy selection
and design.
Squaring the Model with the Data
Model performance is evaluated and improved
by a process of verification, calibration, and
validation (USEPA 2005). Every model should
be carefully checked to ensure that it is based on
accepted scientific principles and that there are
no errors generated by faulty computer code.
This process is called verification.
Models are based on data and scientific
understanding of physical and chemical
processes. Most of the equations in a model
include numerical coefficients. To the extent
that site data are available, some of the
coefficients are based on the fit of the equations
to data, and others are taken to be universal
constants (the acceleration due to gravity being
an example of the latter). Where site-specific
data are limited, coefficients may be values
from scientific literature. Calibration of a model
is the process of adjusting its coefficients to
attain optimal agreement between model-
calculated values and actual site data. Most
commonly, model calibration consists of fine-
tuning the model to provide the best fit to site
data.
The objective of calibration is to make the
model as accurate as possible in its predictions.
This accuracy is further tested through a process
called validation. Normally in validation, a time
period is simulated that is different from the
period that was used to calibrate the model, and
the model is run without changing any of the
coefficients that were adjusted during
calibration. This may require using only a
portion of the data for calibration, thereby
holding data from the remainder of the time
period in reserve for validation. Calculated and
actual values are compared, and if an acceptable
level of agreement is achieved, the model is
considered validated. If not, then further
analysis of the model is performed, leading to
refinements that should improve the accuracy of
the model.
Persistent sediment contaminants can pose very
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
long-term risks. The strong association of
hydrophobic contaminants with flow-driven
particles requires that models must be accurate
with respect to flows and solids movements, in
addition to the behavior of the contaminants
themselves. For these reasons, calibration and
validation datasets should ideally cover long
enough time periods to capture the range of
variability of flow, extreme events causing
erosion, and measureable sedimentation. This is
needed to ensure the model's accuracy with
respect to the build-up and erosion of buried
contaminant deposits. Fitting the model to
shorter time periods and more limited data
targets can give a false impression of predictive
accuracy. For sites with limited historical data
where it has been decided that a mathematical
model is needed, it is important to begin
collecting needed data as soon as possible.
The concepts of model calibration and
validation are illustrated in Figure 3, where a
hypothetical model's development and
application is depicted in terms of modeled
surface-weighted average concentrations
(SWAC), an example of a spatial interpolation
method. The model is initiated with a dataset
describing initial conditions, calibrated to a
dataset spanning a long (e.g. 20 year) period,
and validated against data collected at the end of
the calibration period. The model can then be
used in two predictive capacities, as a
hindcasting tool to simulate how contaminant
levels and exposures likely changed historically
before the calibration period, and as a
forecasting tool to predict future changes to the
system. In this example the model is used to
simulate three remedy alternatives: a Monitored
Natural Recovery (MNR) alternative that
monitors the system as naturally occurring
processes reduce sediment concentrations; a
dredging alternative that causes a short-term
increase in sediment concentrations, followed
by a general decrease and continuing recovery
due to natural processes; and a capping
alternative that reduces surface sediment
concentrations by adding an engineered layer of
clean material.
Uncertainty in the model's prediction of SWAC
is represented by the two sets of dotted lines that
bound the upper and lower range of predicted
concentrations for the two simulated remedies.
The uncertainty bounds are tightest in the
calibration and validation periods, where data
are richest and constraints on the model are the
greatest. The uncertainty increases as the
hindcast and forecast extrapolate further from
the calibration period. The topic of uncertainty
is discussed in greater detail in a following
section.
Models integrate data and scientific knowledge
to better understand the connection between
contaminant releases and risk at a specific site.
Models do not create data, but should be
consistent with available data, for which they
provide a means of synthesis and understanding.
Ideally, model formulation should proceed in
tandem with the site investigation. The CSM
can identify media, processes, and locations of
greatest interest and focus the data collection
effort. This ensures that the resources devoted to
site investigation provide information that is
useful for risk management and remedial
planning. In turn, data from the site
investigation can be used to test the conceptual
model's hypotheses and to improve the
predictive power of a mathematical model.
Models are Approximations with
Specific Objectives
It is important to recognize that all models
simplify complex processes, and that the
objective of modeling is to adequately represent
the processes of greatest importance, rather than
fully describe every aspect of sediment
contamination. (USEPA, 2008a, Glaser and
Bridges, 2007). Just as a good map shows key
features and suppresses unwanted detail, to
highlight the information needed for the map's
intended purpose (compare, for example, a road
map and a weather map), a model includes
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
hindcast
calibration
forecast
i i
Pre-design
Monitoring
Remedial
Activity
Long-term
Prediction
Time
Figure 3: Model Application to a Remedy Evaluation: Calibration, Hindcasting and
Forecasting
(Note: Dotted lines represent 95% confidence intervals for predicted average concentrations)
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
representations of key processes needed for
specific objectives. This makes models
practical tools for problem solving, and at
the same time ensures that they will never fit
data perfectly. In addition, there are real
gaps in scientific understanding of natural
systems, and practical computational
limitations on simulating events at the very
fine space and time scales at which real
processes take place, especially when long-
term forecasts are required. A model should
be thought of as an approximation of reality,
representing the processes that are most
WHAT'S IN A MODEL?
important for making realistic predictions of
exposure and risk over the time frame of
concern. "Models will always be constrained
by computational limitations, assumptions,
and knowledge gaps. They can best be
viewed as tools to help inform decisions
rather than as machines to generate truth or
make decisions." (NRC, 2007a) By using
the best scientific understanding to select
and represent key processes, it is possible to
forecast the future under various remedial
scenarios, and to evaluate relative risks.
Elements of Models
Contaminated sediment sites are dynamic
systems in which exposures can change over
time due to human activities and natural
forces. Contaminants are usually present in
multiple forms such as dissolved,
particulate, and vapor-phase, all of which
have the potential for exposure to receptors.
The contaminants that persist in sediments
are those that tend to adsorb to solids, with
smaller fractions in dissolved and vapor
phases. A model of potential exposures due
to sediment contamination tracks the
contaminant as it is distributed among these
different physical forms, and as it is
transported into, out of, and around the site.
Contaminant transformation and transport is
typically driven by natural processes. For
example, the change in the water column
concentration of an adsorbed contaminant
(i.e. on suspended solids) on any given day
depends on inputs that include any ongoing
contaminant loads, the flow rate, and the
temperature. A model of adsorbed
concentration expresses how contaminant
concentration rises and falls from day to day
as a function of these other variables. In this
example, adsorbed contaminant
concentration is an indicator of the state of
the dynamic system, a state variable. The
concentration of dissolved contaminant in
surface water is another state variable.
Changes in sediment loads, contaminant
loads, and weather force changes in state
variables and are therefore called forcings.
When we express forcings as functions of
time (such as a series of daily flows or
temperatures) they are called forcing
functions. Because contaminant
concentration is dependent on the other
variables, it is called a dependent variable
and the variables that affect it are
independent variables. A summary of some
of the major components of a typical
sediment, contaminant, and fish
bioaccumulation model is presented in
Figure 4.
Complexity and Scale of Models
At the outset when developing a model,
decisions must be made about the degree of
complexity that is justified. Simple models
are more easily understood by scientists,
decision makers, and the public and they
may provide a reasonable degree of
accuracy for minimal investment.
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Complexity can provide added utility, but at
the cost of time and resources, and its utility
depends on the questions asked and the
expected benefit of answering them more
accurately. More complex models require
more effort to construct and more data to fit.
However, if a complex model is not
appropriately constructed and tested,
consistent with management questions and
supporting data, it cannot provide better
answers; it will only be more difficult to
understand, take longer to develop, and cost
more. As noted previously, managing the
complexity of a model and keeping the
focus on model utility is a critical task to be
addressed by a modeling team and their
consultants. The EPA Superfund Sediment
Resource Center (SSRC) provides valuable
guidance and support to project teams.
Floodplain
Deposition
and Erosion
(http://www.epa.gov/superfund/health/
conmedia/sediment/ssrc.htm)
The questions to be addressed using the
model may be relatively qualitative (could
sediment be transported from point A to
point B?) or more quantitative (how much
sediment and associated contaminant is
transported every year from point A to point
B, and what is the trend, if any, in those
amounts?). A remedial project manager
(RPM) should decide whether a qualitative
or quantitative answer is needed, what
project resources are available to support the
decision, and consequently what level of
complexity in analysis or modeling is
appropriate. It is always good practice to ask
"if we add a level of complexity, what (if
anything) can we simulate more accurately,
and how will that improve management
decision-making?"
Water
Upstream
Flow
Upstream
Loading
Particle-bound
Partitioning
Dissolved
Chemical
Mixed
Layer
Buried
Sediment
Resuspension
Particle-bound
Chemical I Partitioning
Downstream
Flow
Downstream
Loading
Dissolved
Chemical
Benthic and
Hydrodynamic
Mixing
Chemical
Transformation
or Biodegradation
Groundwater
Advection
Figure 4: Simplified Summary of Processes and Variables in a Contaminant Transport and
Fish Bioaccumulation Model
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
10
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Dimensions of Model Complexity
Dimensions of complexity include statistical
versus mechanistic, coarse vs. fine', steady-
state vs. dynamic', and deterministic vs.
stochastic. Statistical (sometimes termed
"empirical") models, such as the regression
models discussed above, rely on observed
correlations between variables rather than
mathematical representations of known
physical, chemical, or biological processes.
Statistical (Tier 1) models are simple to
construct and apply, but are suspect if
conditions change relative to the period of
data collection. Mechanistic process-based
(Tier 2-4) models are more burdensome to
construct, but can be more accurately
descriptive, and if processes are correctly
specified they can help to extend forecasting
outside the range of current site data. Most
models employ a mix of statistical and
mechanistic equations, with the split
depending in part on which processes are well
enough understood to formulate
mechanistically.
Choices also need to be made about coarse vs.
fine resolution, in both space and time. Spatial
resolution amounts to the sizing of model grid
cells. In a sediment transport model, cells
usually extend down vertically as well as
covering the site horizontally. Time resolution
is the degree to which the simulation period is
divided into incremental units of time, or time
steps. The forecast for an individual cell and
for a particular time step represents an average
over the whole cell and time interval, even
though there is real variation in each
dimension. Cells and time steps should be fine
enough that forecasts for particularly
important points in space and time are not
lumped with neighboring areas and time
periods. Significantly more input data are
needed to support higher resolution, and
resolution should be increased only as needed,
in a way that improves the value of the model
as a hypothesis testing and remedy evaluation
tool.
Steady state modeling assumes constant
forcings and produces constant values of state
variables, whereas dynamic modeling allows
forcings to vary over time, with resulting
dynamic behavior in state variables. For some
purposes, such as identifying the strength of
relationships between variables, steady-state
modeling can be simple and instructive. For
example, steady state simulations at two
different flow rates, all else held equal, can
produce a clear illustration of the effects of
different flows on sediment transport.
However, variability can be important in
itself. An example is the tendency of recently
deposited sediment to erode during rising
flows. At the tail end of an event, the opposite
occurs, with sediments depositing, and a
steady state simulation could not capture both
phenomena.
Much modeling is deterministic, which means
that a single value is determined for each
dependent variable in each forecast period,
intended as a best estimate. In contrast,
stochastic modeling produces ranges of
forecast values. When inputs are uncertain but
their uncertainty bounds are known, this
information can be used to generate a range of
forecast values. Depending on the project,
there may or may not be value in
understanding this range, and the modeling
team needs to assess the value of adding a
stochastic model. When avoiding worst cases
is of great concern in planning a remedy,
stochastic modeling can be very useful in
estimating the likelihood of those worst cases.
But as with any modeling exercise, the value
of stochastic modeling should be weighed
against the additional cost of developing
stochastic simulations.
The most useful models are not necessarily the
most complex, they are models that are best
designed to answer site management
questions. The best compromise between
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
11
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
simplicity and complexity always depends on
the specific questions asked of the model, the
resources available to build and run it, and the
degree of certainty needed. These factors and
their implications for modeling vary
considerably from one site to the next, and
should be evaluated on a site-specific basis.
Linked Models for Sediment Sites
Models used for sediment management often
comprise linked models of water, sediment,
contaminants, and biota. There are many
sediment and chemical fate and transport
models available today, describing a broad
range of processes, with widely varying
degrees of complexity, and many different
authors and levels of support. Similarly, food
web models are numerous and vary greatly in
complexity and in ease of use. A brief
summary of major categories of models is
provided in this section, and examples are
listed in Appendix A.
Hydrodynamic models of flow use flow
records taken at fixed gaging stations and/or
downstream receiving water levels, both of
which are forcing functions for the model.
Hydrodynamic models take into account the
bathymetry (depth and width) of the water
body, subdivide the waterbody into a grid of
model cells, and route the input flows from
upstream to downstream from one cell to the
next. The purpose is to predict local velocities,
which may increase or decrease depending on
changes in bottom slope and cross-sectional
area. Hydrodynamic models depend primarily
on the physics of water flow, considering the
effects of flow inputs such as watershed flows,
tributaries, groundwater-surface water
interactions, and lake or ocean boundaries,
and taking into account local friction due to
bed roughness, vegetation, or engineered
surfaces. Hydrodynamic models of rivers can
usually be closely calibrated to data on water
levels or stream velocities.
Inputs of solids are carried into the system by
flows, and can deposit as bed sediment at
lower flows and resuspend at higher flows.
The relationship of solids movement to flow is
described by a sediment transport model.
Solids inputs from upstream and from smaller
tributaries are a key forcing function for the
sediment transport model. The sediment
transport model describes how those solids are
transported downstream over time in the form
of suspended load or bed load. Bed load is a
movement of solids skipping and rolling along
the sediment bed, whereas suspended
sediments are distributed by turbulence from
bottom to top of the water column.
Like the hydrodynamic model, the sediment
transport model divides the sediment bed into
cells so that it can make local predictions
about sediment accumulation or erosion. It
may also include vertical layering to predict
changes in sediment bed elevation over time.
A sediment transport model uses the local
velocities provided by the hydrodynamic
model to predict local erosion or
sedimentation of solids. In general,
sedimentation is simulated in models at the
lowest velocities, when velocities are too low
to keep sediment in suspension. Erosion is
simulated in models when the bottom shear
stress created by flow reaches a critical level,
and greater erosion rates occur at higher
velocities. Sediment transport models can be
calibrated to concentrations of suspended
sediment and to measured erosion and bed
sedimentation rates.
The movement of sandy sediments is better
understood scientifically and easier to predict
with equations than the movement of finer
particles like silts and clays. This is because
the cohesiveness of those smaller particles is a
complicating factor and differs from site to
site. The movement of sand particles in
response to flows can often be described by
standard textbook equations that are
applicable at any site. For silts and clays,
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
12
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
however, the tendency to erode depends on
complex characteristics of the local sediment,
and has not been fully explained in terms of
simple, measurable properties (like bulk
density and grain size). Consequently, the
equations that models use to describe erosion
are often based on direct measurements of
local erosion. Ideally, these measurements are
made with flumes placed on the sediment bed
or operated in a laboratory, applying a flow to
erode sediment samples collected at the site. If
flume measurements are not available,
standard erodibility equations can be
calibrated to data on suspended sediment
concentrations measured at varying stream
velocities, but are subject to greater
uncertainty in forecasting. Models of erosion
of cohesive sediment also take into account
the age of sediments, such that fresh
sediments are predicted to erode more easily
than sediments that have had days or weeks to
consolidate.
The hydrodynamic and sediment transport
model components set the stage for modeling
of contaminant transport. Contaminants can
be dissolved, associated with solids, or can
occur in the form of a gas. They can also be
associated with colloids, which are suspended
solids that are so fine that they do not settle at
the lowest velocities. Contaminant transport
models include a partitioning component that
determines how much of the contaminant
appears in each form in the sediment bed,
including the pore water that surrounds
sediment grains, in groundwater that may pass
through the sediment bed, and in the overlying
water column. The bed and water body are
divided into a grid of model cells, and a set of
partitioning calculations is performed to
distribute the contaminant in each cell into its
various forms. Partitioning calculations are
usually based on published partition
coefficients that describe a chemical's
tendency to adsorb to solids, dissolve, or occur
as a gas. These coefficients are chemical-
specific, but the hydrophobic organic
chemicals that are the targets of many
sediment clean-ups are primarily associated
with solids, especially with fine-grained
material having a high organic carbon and/or
black carbon (e.g., soot and char) content. The
relationships used by models to describe those
strong associations between organics and
solids depend on the organic carbon partition
coefficient (Koc) and on the organic carbon
content of solids. Predicted concentrations in
each form and location can be compared to
actual contaminant samples from those media,
and the model calibrated to more closely
predict those concentrations.
Site-specific measurements of partition
coefficients should be used if they are
available. If not, which is often the case, a
comprehensive handbook of chemical-specific
partition coefficients, such as Kow, Koc, and
Henry's Law Constant (H) is Mackay, et al.,
(1992), now available in a continually updated
CD-ROM from CRC press. EPA also supports
a set of software products, called EPI Suite,
that can be used for estimating many of the
chemical-specific parameters used in models
if site-specific measurements are not available
(http://www.epa.gOv/opptintr/exposure/pubs/e
pisuite.htm).
Often mixtures of chemicals, such as total
PCBs or PCB homologs, are modeled. In this
case, it is most desirable to obtain site-specific
partitioning measurements, followed by minor
adjustment during model calibration (Bierman
etal., 1992; Butcher et al., 1997). If that
approach is not possible, a partition coefficient
value for these mixtures may be estimated by
weighted-averaging of the literature values for
the individual chemicals found in the mixture.
Models of metals contamination also use
equations to represent partitioning to solids
versus the dissolved form. Partitioning of
metals to solids can depend very strongly on
oxidation-reduction conditions in the
sediment. In oxygen-deprived sediments,
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
13
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
sulfur appears in the form of sulfide and bonds
with metals to form insoluble sulfide
compounds. Where oxygen is more abundant,
sulfide is transformed to sulfate and metal ions
are released and can occur in dissolved form.
For both organics and metals, partitioning is a
key to bioavailability. Sediment-dwelling
organisms, for example, may be much more
vulnerable to pore water contamination than to
contamination that is tightly bound to
sediment particles. Partitioning also affects
transport from the source area to downstream
and downwind locations. This includes
dissolved contaminant traveling downstream
at a rate that is predicted by modeled water
flow, and contaminant that is adsorbed to
solids, moving with those solids as they settle,
resuspend, and travel in suspended form.
In some cases, particularly low- flow
conditions, groundwater may be a pathway for
transport of contaminants through the
sediment bed. While groundwater models
may be of importance in characterizing
groundwater-surf ace water interactions in
sediment models, their use is not detailed in
this document. An additional process that is
usually modeled as a loss to the aquatic
system is contaminant in vapor form escaping
to the atmosphere and transported away from
the site. At some sites, air is a significant
exposure pathway, and in such cases
concentrations in ambient air should also be
simulated.
While direct exposure to contaminants may
pose risk, the most important risks to human
and ecological health can arise through
bioaccumulation. Food web models translate
the dissolved and adsorbed concentrations
provided by the contaminant transport model
into body burdens for target ecological
receptors, including fish and wildlife species,
and the species that comprise the food chain
that sustains their populations. (As an
example, Figure 5 shows results from a food
web model developed for PCBs in the Lake
Ontario ecosystem.) Food web models
represent populations of birds, fish, and their
prey by age cohort and area. The
accumulation of contaminants in organisms is
essentially the difference between
contaminants accumulated via food and water
and contaminants lost via respiration and
excretion.
In food web models, the contaminant is
bioaccumulated as a byproduct of obtaining
energy through food, and the fraction
absorbed by the body is governed by a
partitioning equation, similar to the
partitioning between water and solid particles.
Partitioning to tissue depends on the fat (lipid)
content of the organism, just as solids
partitioning depends on organic carbon
content (see more detail on page 18). Some
food web models also simulate foraging as
movement between available habitats in
response to their characteristics and
availability of food. "Static" applications of
foodweb models depict organism contaminant
concentrations at a point in time under
specified conditions. "Dynamic" applications
simulate changing concentrations in the
organisms over time. As individuals are
simulated to age and grow, they build up
contaminant in their tissue. In this way, food
web models represent the accumulation of
contaminant over the life cycle of the
organism, simulating real-life relationships
between age, size, and location of biota and
their contaminant body burdens.
With data on current conditions and the
combination of hydrodynamic, sediment
transport, contaminant transport, and food web
models, future tissue concentrations for target
species can be forecasted. A forecast
simulation begins with a set of initial
conditions, including estimates of current
contaminant concentrations in sediment,
water, and biota.
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
14
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
fake trout
770
" each dot represent* "available" chemical water concentration -O.OOOlSpph
:::^
:::::I[[[i::::::::::::::::::::;;
Figure 5: Bioaccumulation of PCB in Lake Ontario
With linked models, initial conditions are
often defined at different spatial scales. A
river reach may have many model cells, with
the initial conditions in each cell based on
sediment samples obtained within that cell,
whereas a food web model may represent the
same reach as a single model cell, with initial
conditions based on fish sampled throughout
that reach.
Additional inputs required as forcing functions
include long-term daily series of any ongoing
contaminant loads, based on best judgment
about future source control, and of future
solids loads, flows, and temperatures.
Assumed solids loads will depend on
assumptions about future watershed
development and management, and flows and
temperatures are derived from the historical
record and can be adjusted to reflect any
expected changes in conditions.
The model proceeds from one time step to the
next, routing stream flows through the system,
moving sediment, partitioning and
transporting the contaminant downstream, and
simulating biological uptake of contaminant.
The result of the forecast is a prediction of the
spatial pattern of contaminant exposure and
bioaccumulation as it develops over time,
reflecting the effects of legacy contamination
and changing conditions. This may include
increasing exposure due to erosion and/or
decreasing exposure due to burial of the
contaminant.
When this forecast of exposure is combined
with other exposure and effects data, such as
frequency of fish consumption and health risk
factors associated with the contaminant, the
result is a forecast of risk as it can be expected
to change over time, i.e., under the no-action
and MNR alternatives.
To evaluate the potential benefits of
remediation, a similar forecast can be
generated with updated bed conditions,
reflecting concentrations that are expected to
be present after a remedial action. The
difference in the two forecast outcomes, with
and without changes in assumed bed
conditions due to remedial action, represents
the expected net benefit of remediation. The
time pattern of those predicted benefits may
United States
Environmental Protection
Agency
Office of Superfund Remediation and
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
reflect tradeoffs between present and future: addition to removal of some of the
for example, a dredging remedy is likely to contaminant mass. The use of models to
cause some sediment resuspension and predict remedial outcomes is described in
contaminant release and transport initially, but greater detail in a following section.
may lower future exposures and risks, in
Process Representation in Models
Models provide a framework by which many different physical processes can be identified,
quantified, and compared in terms of their relevance to a particular risk management
endpoint. For example, processes like particle settling and deposition, event-based
resuspension, and bioturbation may all be important in controlling how contaminant levels in
surficial sediments change with time. Or, if a goal at a particular site is to limit present and
future exposure of a particular fish species to sediment bed contaminants, models can
assess the processes that control that exposure, and identify which processes or parameters
actually matter most. The example below highlights how model development and revision
can inform exposure processes, risk assessment, and remedy selection.
Example:
Problem: In a slow-moving river in a mostly agricultural watershed, sediments downstream
of a historical chemical waste recycling operation show elevated levels of PCB at 3-4 feet of
depth, and lower contaminant levels at shallower depths. Measurements of the age of
sediments at depth were made using geochronological dating methods. MNR is proposed as
a remedy alternative to limit exposure of benthic feeders to surficial sediments.
Model Application. An Environmental Fluid Dynamics Code (EFDC) model is constructed
using measurements of upstream suspended sediment concentrations across a range of
river flow rates to provide an estimate of solids load to the upper river. A problem arises
during model calibration: when the model is calibrated to present-day suspended sediment
levels throughout the river, the predicted rate of sediment deposition and burial isn't high
enough to explain the depth of burial of the contaminants, given what's known about
historical contaminant releases.
Conceptual Model Revision: Further exploration of the upstream solids load shows that
historically, solids loads were much higher because there were fewer agricultural runoff
controls in the 70's and 80's. A revised model that incorporates a long-term trend of
decrease in upstream solids load is used to reproduce historical deposition and forecast the
future rate of burial of bed contaminants.
Outcome: The model shows that the sediment recovery rates predicted by geochronological
cores probably overestimate the present-day rate of recovery of the system, due to the
expected continuing decrease in upstream sediment supply and its effect on sediment
deposition in the vicinity of the PCB deposits. The MNR assessment is adjusted to reflect a
slower rate of recovery than initially envisioned.
United States Office of Superfund Remediation and OSWER Directive 9200.1 -96FS
Environmental Protection Technology Innovation November 2009
Agency
16
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
WHAT'S NEEDED FOR A MODEL?
When a site is listed on the National Priorities
List (NPL), there is typically some
information and data on the extent of
contamination. Sources may have been
identified, and specific ongoing migration
pathways may be suspected. To refine this
thinking into a solid CSM, and to build a
process-based numerical model if appropriate,
additional data are needed. Ideally, those data
can be used to quantify any continuing sources
of contamination; tabulate the range of water
flows in the contaminated area; show how
solids move between the sediment bed and a
suspended state at various flows; determine
how much of the contaminant is dissolved and
how much is attached to solids; and show how
much contaminant is taken up by target fish
and wildlife species through their life cycle.
Through data collection, each link in the CSM
is tested and verified or refined.
Data collection is costly and its purpose is not
to be exhaustively descriptive of the site. It
should be carefully targeted to the key
sources, pathways, and exposures in the CSM,
especially those that are key site-specific
determinants of human health and ecological
risks. If this is done effectively, then the data
set will support model construction, and
model development will also assist with
identifying data needs. On the other hand, if
there is not enough thought given to these
conceptual links in the data collection phase,
then the task of predicting the effects of
remedial management actions will be much
more difficult, whether or not this is done with
a numerical model.
Sediment Data
Modeling begins by quantifying current
contaminant inventories in sediment, and
other sediment characteristics that can affect
concentrations and sediment stability. These
are the model's initial conditions. The
following is a representative set of data to
collect on sediment, described in greater detail
below:
Contaminant concentrations
Organic carbon content
Acid volatile sulfide (AVS, when metals
are present at concentrations that may pose
risks)
Dry bulk density (g/cm3 dry weight)
Grain size distribution
Sediment sites can be vast, and sampling is
costly, so the sampling plan should be
carefully designed to provide good coverage,
both horizontally and vertically, of
contaminant deposits, while collecting only
the data needed to support a risk management
decision at the site. Sediment core samples are
preferred to surface grabs because it is
important to include subsurface sediment
characteristics, including contaminant
concentrations, in models. In the vertical,
cores should be segmented in such a way that
vertical layering and lower bounds of deposits
can be identified. This depth could be based
on an estimate of the historical rate of burial
(such as from navigational dredging records)
and knowledge of the dates of release.
Judicious analysis of a subset of core
segments and archiving of others can
minimize the analysis of uncontaminated
samples at depth.
In the horizontal, samples should be
distributed so as to try to capture
concentration peaks and trends. It is a fact of
life for large sites that sediment core data will
be used to represent large areas between cores,
with uncertainty about sediment
concentrations at unsampled intermediate
locations.
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
17
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Sediment organic carbon content is important
to analyze because of the chemical affinity
between it and nonpolar organic chemicals.
This strong affinity makes organic
contaminants less bioavailable, thus
decreasing contaminant bioaccumulation in
organisms. The situation is similar for metals
in sediments with high sulfide content. In this
case, metals tend to be present in insoluble
form, and are much more difficult for
organisms to assimilate in tissue than in
dissolved form. These differences in
bioavailabilty are important from a risk
perspective.
Bulk densities and grain sizes are important in
predicting erodibility. Sediment with lower
bulk density is generally more easily eroded.
For grain size, coarse particles (sands and
gravels) are less erodible than finer particles.
Fine-grained sediment, which includes a high
proportion of silt and clay, forms a cohesive
bed and erodibility is highly site-specific, as
discussed on pages 12-13.
Hydrodynamic Data
Data required for hydrodynamic models
include inflows, physical boundaries of the
water body, and any water levels that can be
considered fixed for purposes of the model.
Typical river flows, for example, depend on
flows from upstream and from tributaries, the
geometry of the channel, and the water surface
elevation of a downstream water body. The
most basic predictions of these models are
velocities and water levels, so those data are
needed as targets for calibration. Thus, the
dataset to build a hydrodynamic model will
typically include:
Bathymetry and shoreline geometry
Upstream flows, preferably from a reliable
gage with a lengthy historical record
Watershed drainage areas for important
ungaged tributaries
Water levels at any downstream
boundaries (e.g. river, lake, or tidal
boundary), preferably from a reliable gage
with lengthy historical record
Stream velocities
Water surface elevations
Bathymetric information and shoreline
geometry can be collected at a very fine scale,
but the degree of hydrodynamic model
refinement should still be governed by the
anticipated use of the model. For example, if it
is important to estimate stream velocities and
associated erosive forces for local site features
of a specific size, to support remedial
planning, then the grid in that area should be
refined. However, an overly fine
hydrodynamic grid could cause model run-
times to be excessive without adding
significant accuracy to forecasts of sediment
and contaminant transport.
Data on Solids. Erosion, and
Deposition
Data required for sediment transport models
include information on the solids in the
sediment bed, as initial conditions, and solids
in suspension and moving along the bed as
bed load, as calibration and validation targets.
It is important to understand the movement of
sediment under normal day-to-day conditions
and during extreme events. Outputs of the
hydrodynamic model, including water depth,
velocity, and shear stress, are also inputs to
the sediment transport model, Suspended
solids and rates of bed erosion and
sedimentation are outputs of the sediment
transport model that can be compared to data
for model calibration and validation. A basic
dataset to build a sediment transport model
will typically include:
Water column samples of suspended
solids, sampled over a range of flows
Bed load flux rates and physical
properties, if bed load is present
Flume studies of erodibility at high
velocities and shear stresses
Long-term measures of erosion/deposition
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
18
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Water column suspended solids data should be
collected at enough stations on the water body
to capture any longitudinal changes in
suspended solids due to erosional or
depositional areas. Bed load, if present, can be
quite local, and bed load samplers should be
deployed along transects with close spacing.
Flume studies can be used to provide a
measure of the erodibility of sediments. They
are generally unnecessary with noncohesive
sediments (sands and gravels), whose
properties are well-known, but are valuable
with cohesive sediments (silts and clays)
because their erodibilities vary considerably
from site to site.
In contrast, long-term erosion/deposition data
provide an integrated look at changes in the
sediment bed over the full range of high and
low flows that have occurred. These rates can
be inferred from successive bathymetric
studies, navigational dredging records,
sediment traps, erosion pins or chains, and
analysis of vertical distributions of certain
radioisotopes (notably Cs-137 and Pb-210) in
sediment cores. These measurements can be
highly variable from year to year and location
to location, so the use of several types of data
from several locations is strongly encouraged.
Flume studies and measurements of long-term
erosion and deposition should employ samples
and data from enough locations to provide a
representative picture of sediment erodibility
for distinct areas of the site. This includes
uncontaminated as well as contaminated areas,
because sediments eroded from
uncontaminated areas can contribute to burial
of contaminants in downstream portions of the
site.
For eutrophied waterbodies, algal solids can
serve as another solid transport medium for
solids. A model of algal solids growth is an
important piece of the overall solids balance in
these cases. These models are complex, and
require data on nutrients and sunlight as
inputs, and chlorophyll as a calibration target.
It is important to stress that the data required
to support an extensive sediment transport
modeling effort can be difficult and expensive
to collect. Sediment transport is often strongly
impacted by high flow events, which can be
very difficult to monitor and often produce
highly transient data that can be challenging to
interpret. Sediment characteristics that affect
transport are typically highly heterogeneous,
and that heterogeneity can drive extensive
data collection efforts if not limited by careful
consideration of the real needs of the project
and the modeling required to support it. For
sediment transport modeling in particular, it is
critically important that the project team
carefully consider the level of model
complexity required to answer project
questions, and the corresponding degree of
data richness and resolution required to
support model development.
Contaminant Data
Hydrodynamic and sediment transport models
predict the downstream movements of water
and solids. A contaminant transport model
builds on those predictions to simulate the
transport and environmental fate of the
contaminant. The initial conditions describe
the initial mass, extent, and distribution of the
contaminant. Its predicted fate and transport
is then governed by partitioning to water and
solids (i.e., to particulate organic matter and
inorganic sediment) as they move through the
system. Snapshots of adsorbed and dissolved
contaminant concentrations at several
locations in the water body on sampling dates
that span a representative range of flows serve
as calibration targets. Additional data required
for contaminant transport models include:
Partitioning coefficients for chemicals of
concern (typically from literature or
handbooks), including Koc and the Henry's
constant governing volatilization
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
19
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Dissolved, colloidal, and adsorbed
concentrations of contaminants in the
water column, sampled over a range of
flows and temperatures
Like suspended sediment samples, water
column contaminant samples should be
collected at enough stations to capture
longitudinal changes due to erosion and
deposition. In cases where pore water plays a
significant role as an exposure or transport
pathway, pore water sampling may be a useful
additional source of information (USGS,
1998).
Biological Data
Food web models build on predictions of
flows and contaminant concentrations, adding
contaminant uptake by local biota. To do so
the food web model needs to simulate local
populations of representative food web
species, including their dietary preferences
and contaminant exposures. Tissue
concentrations are the calibration targets for
these models. Typical data for food web
models include:
Identification of endpoint species, diet,
and predator/prey relationships
Tissue contaminant concentrations, along
with data on model inputs including age,
size, and lipid, moisture, and solid
contents of organisms
Temperature and dissolved oxygen
concentrations
Biota are exposed throughout their home and
feeding ranges, so sampling should target
species that are not exposed to contamination
outside the site that is to be remediated, to
give a clear indication of local exposure and
uptake. Tissue concentrations can be highly
variable, and numbers of organisms sampled
should be planned carefully to ensure that
their medians or averages are representative of
the population, with acceptable standard
errors. If data are collected in successive
rounds, variances from initial data can be used
in planning sample sizes for subsequent
rounds (USEPA, 2008b).
It is worth reiterating that data collection is
costly, and no dataset can form an exhaustive
description of the entire system. Data
collection resources should be focused on the
sources, pathways, and/or exposures that are
most important from the standpoint of risk
reduction, and that are least well understood.
Data collection is justified to the extent that it
can help in understanding these relationships,
especially as that understanding informs
remedial decision-making.
Single Versus Multiple Rounds of Data
Important sediment and contaminant
processes take place over a range of time
scales. Stream flows vary daily, weather
conditions vary seasonally, and long-term
changes in the sediment bed take place over
periods of years. Depending on the
requirements of the investigation, data
collection activities may focus on short-term
fluctuations, seasonal changes, or long-term
trends.
Short-term data are important for
understanding variability in exposures,
especially changes due to flows, including
extreme flow events. A model based on short-
term data is best suited to simulating the near-
term range of potential exposures, given
expected frequent variations in flow and
temperature, assuming implementation of
alternative remedies. Data collected over
multiple years are essential to making realistic
predictions of long-term trends in sediment
and tissue concentrations. This understanding
of trends is especially important in predicting
the long-term effects of remediation, because
all remedies depend to some extent on natural
recovery processes, and these processes take
time. Collecting multiple rounds of data on
concentrations in surface sediments, water,
and biota, from the commencement to the
completion of the site investigation, provides
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
20
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
the clearest picture of long-term trends, which of the time needed to achieve long-term
can serve as long-term calibration targets for remedial targets (Figure 3).
models and improve the accuracy of forecasts
Models and Data Collection
Everyone knows that models need data, but what's less well known is that data collectors need
models. An initial site investigation that is guided by an accurate conceptual understanding of
system behavior (a conceptual model) is likely to be far more efficient than a purely exploratory
data gathering effort. And, a site investigation that builds on previous phases of investigation
and model development can be highly targeted, making measurements only where needed, as
guided by the model. Models provide a framework for organizing data, and for testing
consistency with other sources of data and physical processes that govern site behavior. And
in turn, collected data provides the basis for further model development and refinement. The
example below illustrates how data gathering and model development can proceed hand-in-
hand at a contaminated sediment site.
Example.
Background: A long-term dataset of fish contaminant levels collected at a contaminated
sediment site shows a long-term trend of decreasing body burdens with time. In order to better
understand the relationship between fish contaminant levels and sediment concentrations, a
model is developed in the Gobas framework, drawing on the existing dataset of surficial
sediment contaminant levels, fish body burdens, and a limited set of benthic invertebrate data.
A parallel laboratory investigation is conducted to develop biota-sediment accumulation factors
(BSAFs) for bottom-dwelling fish using site-specific sediment samples.
Model Application: Predicted fish body burdens, based on a Gobas model calibrated to fish
and sediment data from the site, indicate less bioaccumulation than predicted by the
laboratory-based BSAFs. The generally lower site fish body burdens suggest that the mode of
exposure to sediments may be affected by spatial variability in contaminant levels and fish
habitat preferences, making it worthwhile to account for differential exposures by habitat type in
the food web modeling.
Additional Data Collection: A sidescan sonar survey of the sediment bed shows significant
variation in the texture of the bottom substrate and the associated habitat quality. A series of
electroshocking surveys exposes a strong fish preference for irregular bottom structure,
particularly rocky substrate where contaminant exposure is lower than in broad areas of
sediment where deposited fines have high levels of contaminants. Incorporation of these
findings into aversion of the Gobas model that accounts for such preferences results in a
reduction of scatter in model calibrations, improved predictive capacity of the model, and
reconciliation of model results with laboratory BSAF studies.
Outcome: The strong habitat preference expressed by the target species results in a
reassessment of remedies, with a greater emphasis placed on habitat improvement paired with
reductions in exposure concentrations.
United States Office of Superfund Remediation and OSWER Directive 9200.1 -96FS
Environmental Protection Technology Innovation November 2009
Agency
21
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
HOW CERTAIN ARE MODEL PREDICTIONS?
Models make predictions with uncertainty for
a number of distinct reasons. This section
will catalog those sources of uncertainty, and
discuss how to quantify their effects on
forecasts, as well as what level of uncertainty
is reasonable to expect.
Sources of Uncertainty
One category of uncertainty is model
uncertainty: the equations of a model may
not fit the true physical, chemical, or
biological relationships exactly. For example,
modelers often use standard functional forms
for convenience, such as linear or logarithmic
equations, when the true relationship may not
fit either of those functions perfectly. Model
uncertainty can also arise from application of
a model that was calibrated and validated for
a different site, without sufficient
recalibration and validation to adjust for local
conditions. When simulating the effects of a
remedy, data from a pilot study can provide a
basis for recalibration using new initial
conditions. This can reduce model
uncertainty.
Input uncertainty is a second distinct
component of uncertainty. Even if the
selected model is correct in form, there may
be errors in model parameters. For example,
coefficients for partitioning between water
and organic carbon can be obtained from
handbooks of chemical parameters, and may
be presented as ranges. A single value must
be chosen, and the uncertainty in this
parameter imparts an uncertainty to predicted
partitioning and subsequent results. Model
inputs, such as initial sediment contaminant
concentrations and flows, are also subject to
error, inherent with the instruments used to
measure them and the reliability and
consistency with which operators make those
measurements. Finally, point estimates of
concentrations and other inputs are averaged
over model cells and treated as representing
the entire cell uniformly. If cell resolution
captures spatial trends well, then this
aggregation error is of little consequence. If
not, then it can be reduced by reducing cell
size and taking enough samples to
characterize each cell. However, such
practices impart additional time and resource
costs, and are justified only to the extent that
the added detail helps inform the selection of
a remedial alternative.
Even if the form of a model matches the
physical phenomenon very well, parameters
are calibrated closely, and inputs are
measured with great accuracy, some
stochastic variability will remain. This
represents factors that cause actual values to
be different from predictions and are not
represented in the model. This may be
because they are not well understood or
because data have not been collected to
explain them. The variability in fish tissue
concentrations of contaminants is a good
example. Much, but not all, of the variability
in fish tissue concentrations can be explained
by the species, age, length, lipid content, and
sediment and water column contaminant
concentrations. The rest of the variation may
be due to natural variability between
individual fish in habitat, diet, or genetic
makeup. Models can predict average fish
tissue concentrations for a given species as
well as the range of natural variability around
that average.
An understanding of all of the above sources
of uncertainty is important for estimating and
controlling the uncertainty in model
predictions of present and future conditions.
Model uncertainty can be reduced by going
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
22
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
through the process of model development,
hypothesis testing, and refinement described
in the previous section on "What's in a
model?" Uncertainty in inputs is translated
into uncertainty in model outputs, and may
be exaggerated or minimized depending on
the sensitivity of the model to those inputs.
Model sensitivity is often measured by
varying inputs by a fixed amount (commonly
10%) and observing the corresponding
changes in key model outputs. This analysis
helps users of models identify the key
parameters driving the uncertainty in model
predictions, and helps define where efforts to
improve predictions are best placed (i.e.,
decreasing the uncertainty of the parameters
to which the model is most sensitive).
Ultimately, the true "rightness" or
"wrongness" (i.e., accuracy) of a model can
never be known perfectly. But estimates of
model uncertainty are possible, and can be
helpful. The following section describes
methods commonly used for quantifying
model uncertainty, and some realistic
expectations and limitations on the use of
models.
How to Quantify Uncertainty
There are two distinct purposes to
quantifying the uncertainty of a model. The
first is to determine whether this model or an
alternate model provides the better fit to the
available data. The second is to estimate
uncertainty bounds on predictions of the
future.
To evaluate the goodness of fit to calibration
data, measures of the typical deviation
between predicted and actual values are
usually used. These can be averages of
deviations, or medians in cases where a few
very large values might distort the average.
Another common measure is the mean of
squared deviations, called the root mean
squared error. This measure is attractive
because it is always positive and provides an
estimate of the typical deviation in the same
units as the variable itself (as opposed to
those units squared.) As an intermediate step
in model development, it is common to look
at a set of these measures of typical model-
data error and choose a candidate model
formulation that performs best for the
majority of the variables and goodness of fit
metrics.
When a prediction is generated from a model,
such as a time series of predicted future fish
tissue concentrations, it is also important to
estimate the upper and lower bounds of those
estimates. In this example, the question may
be whether bio accumulation of the
contaminant might be much lower or higher
than the best estimate. This has traditionally
been done in a number of ways: using
parameter bracketing analyses; Monte Carlo
analyses; by exploring alternative
calibrations; and by making comparisons to
validation data.
In a bracketing analysis, each key parameter
of the model is selected and reset to the top
and then to the bottom of its reasonable
range, based on the judgment of the modeler.
The two new forecasts that result bound the
forecast, in terms of uncertainty in that
parameter. For example, coefficients
governing the adsorption of chemicals to
organic carbon in sediment solids are
measured in the laboratory, and handbook
estimates may not be a perfect match to
adsorption in the field. The type of carbon is
also important: the amount of "black carbon"
is usually unknown, and is a stronger
adsorbent than other forms of carbon, tending
to reduce contaminant bioavailability. A
bracketing analysis shows how much this
uncertainty matters, and how the forecast
could differ with an adjustment in the value
of this one uncertain parameter.
This procedure is sometimes extended to
allow for simultaneous variation in multiple
parameters. For example, a set of parameters
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
23
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
could all be varied within reasonable ranges
in a direction that lowers the series of
predicted fish tissue concentrations to a lower
bound, and then all varied in the opposite
direction to generate upper bound predictions
of future concentrations. This procedure is
not recommended, because it generates best
and worst cases that may be very unlikely
and inconsistent with the data to which the
model was calibrated, giving a distorted view
of the spread between best and worst cases.
A preferred method is suggested by the
calibration process itself. When a model has
been well calibrated, and a parameter is then
adjusted, the fit to data is worsened and
compensating changes in another parameter
or parameters are needed to restore a
reasonable fit between model and data. The
model is then run under a set of alternative
calibrations, each perturbing one parameter
from its calibrated value and then adjusting
other parameters to produce a second-best
model-data fit. The range of predictions
generated by these alternative calibrations is
consistent with the calibration dataset to
nearly the same extent as the calibrated
model, and should therefore be considered to
be within reasonable uncertainty bounds. An
exploration of the possibility of alternate
calibrations of the model and an
understanding of the range in outputs that can
exist among these calibrations provides
important insight into the uncertainty of
model predictions.
A third technique for generating uncertainty
bounds is Monte Carlo simulation. Named
after the casino resort where dice are rolled
and roulette wheels spun in games of chance,
Monte Carlo requires the user to specify a
distribution for each parameter to be varied,
including correlations between parameters
that are not considered to be independent. A
series of forecasts is then generated, each one
based on a draw of each parameter from its
distribution of possible values. From this set
of forecasts, a range of predictions of each
variable at each future date is produced. The
validity of this range depends on the validity
of the assumed parameter distributions,
including correlations between parameters.
The modeler's experience that a change in
one calibrated parameter must be offset by a
change in another to restore goodness of fit
suggests that parameters are typically
correlated, and unless those correlations are
known and specified, the output of a Monte
Carlo simulation may overstate the spread
between extreme predictions. This method is
clearly only practical if multiple model runs
can be generated in a short time, which is not
the case for the most linked
sediment/contaminant transport models.
Finally, an important check on the actual
performance of a model is by comparing to a
dataset independent of the data used to
calibrate. A validation dataset provides a way
to verify the accuracy of model predictions,
identify unanticipated bias in important
outputs, and test the validity of uncertainty
estimates made using other methods.
Realistic Expectations and
Limitations
An important objective of modeling is to
make sound, reasonable predictions. A good
calibration demonstrates that a model can do
this, by simulating state variables accurately
within the calibration period. No model can
fit the data perfectly, however, for the
reasons discussed above. For future periods,
outside the calibration dataset, additional
unknowns may come into play, so differences
between actual and predicted values for
future periods can be expected to be at least
as great as those for the calibration period.
This is why it is good practice to validate a
model for a dataset that is held in reserve
during calibration. The differences between
actual and predicted values in calibration and
validation provide the best indication of the
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
24
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
size of the prediction errors that we can
expect if we use the model to make short-
term predictions.
For contaminated sediment sites, recovery
may take decades, even with active
remediation, so models are used to make very
long term predictions under a variety of
remedial alternatives. Observers of weather
forecasts and economic forecasts know that
uncertainty tends to increase as the period of
the forecast is extended into the future
(Figure 3). This increase in uncertainty is
expected in environmental fate and transport
forecasting for several reasons. For one thing,
there may be unexpected changes in or
shocks to the system that are not reflected in
a model fit to current data. Examples are
possible temperature or flow changes due to
gradual global climate change, flow changes
due to dam removal and/or urban
development, or changes in background
contaminant loads.
In addition, long-term sediment forecasts
predict trends in state variables, and these can
depend on predicted rates of long-term
release or burial of the sediment inventory of
contaminant. If those predicted rates are too
high or too low, then absolute prediction
errors in sediment, water column, and tissue
concentrations tend to grow with time. For a
good long-term forecast, the expectation
should be that predicted rates of growth or
decay in concentrations are accurate, while
the absolute magnitudes of predicted
concentrations are subject to increasing
uncertainty over time. The more thoroughly a
model has been calibrated and validated,
using a data set spanning a number of years,
and the more accurately remedies have been
represented in the model, the more
confidently those trends can be compared.
Comparisons of the predicted absolute
concentrations to target values are more
uncertain, especially if they are the results of
very long forecasts.
Forecasts of concentration trends under
competing remedial alternatives are
discussed in the next section.
Characterization of Uncertainty in Models
Both developers and users of models generally agree that all models are uncertain.
Increasingly, modelers and project managers are moving beyond an acknowledgement of
uncertainty to something more useful - a quantification of the degree of uncertainty in
models, and estimation of its importance in model application. Quantified uncertainty is useful
in designing sampling programs, in answering management questions that have "gray areas",
and especially in remedial decision making. The example below describes the development
of a complex model, and a process by which the major sources of uncertainty in the model
were identified, and impacts on the uncertainty in model output were quantified.
Example;
Background. At a riverine contaminated sediment site, an historical accumulation of
sediments in depositional dam impoundments is now an ongoing source of contaminants to
the river, due to removal of the dams and incising of the river into the former impoundment
sediments. Because the river channel dropped, the river banks upstream of the former dams
are now composed of contaminated sediments that act as a continuing source of solids and
solids-associated contaminants to the river. (cont'd)
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
25
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Model development: An EFDC model is constructed that represents the current distribution
of contaminants in the river bed, banks, and former impoundment areas. Extensive data are
available on historical trends in sediment and fish tissue contaminant levels. An important
input to the model is an empirical analysis of the rate of bank erosion in the former dam
impoundments, and the corresponding contribution of solids and contaminant load to the
river. The river is also impacted by significant flow rate variation, with different mass transfer
processes that operate under low flow conditions (contaminant diffusion from the bed) and
high flow conditions (particle resuspension). The model is developed over several field
seasons to provide a robust dataset and to account for all processes relevant to contaminant
transport, producing a model that is well-calibrated and validated to sediment and
contaminant data, capturing spatial variation in contaminant concentrations and trends in
contaminant levels.
Uncertainty Analysis: Following model development, the modeling team and the project
management team meet to post-analyze the model, trying to understand 1) what are the
most important measures of model performance, and 2) what are the most important model
inputs (processes and parameters) that affect those measures. As critical performance
metrics for demonstrating model success, the group identifies a close correspondence of
model predictions to:
Field measurements of total suspended solids (TSS)
Field measurements of dissolved, suspended and sediment PCB
Site-specific burial rates indicated by geochronological studies
Long-term recovery rates in sediment and fish
While many parameters and processes are identified as important, five are identified as
critical for meeting the above performance metrics:
Flow-dependent particle settling velocity
Sediment resuspension rate coefficients
Critical shear stresses for initiating bank erosion
PCB mass transfer rate coefficients, for example the rate at which PCBs migrate from
sediment pore water to overlying water due to diffusion or groundwater upwelling
Bank erosion rates
In order to explore overall uncertainty in model predictions, the above coefficients are varied
to explore the degree to which they impact model output. The allowable degree of variation
in the above parameters is bounded on both the input and output sides of the model. On the
input side, parameter variation is limited to the known uncertainty in the parameters (range of
reasonableness), and on the output side, the degree to which parameter variation impacts
the quality of the model calibration (calibration constraint). The effect of this range of
variation on long-term forecasts of system recovery provides a useful, meaningful bound on
prediction uncertainty.
United States Office of Superfund Remediation and OSWER Directive 9200.1 -96FS
Environmental Protection Technology Innovation November 2009
Agency
26
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
HOW DO WE PREDICT REMEDIAL OUTCOMES?
Models are often used in Feasibility Studies
to evaluate the short-term and long-term
effectiveness of several alternatives. They
can predict the time path of future exposures
under each alternative, so that likely risk
reduction can be estimated. Risk reduction is
estimated relative to monitored natural
recovery (MNR), which includes monitoring
without active remediation. Calibration and
validation of the model to pre-remedial data
optimizes the representation of natural
processes, which are relied upon by the
MNR alternative. This section explains how
an MNR forecast is developed, and how its
initial conditions are modified to generate
forecasts for active remedies. An MNR
simulation is straightforward to develop,
because it represents a continuation of
natural conditions, and is a natural baseline
scenario upon which to build active remedy
scenarios. The development of updated bed
conditions for representing the expected
outcomes of active remedies using models
and/or pilot studies also deserves
considerable effort to minimize uncertainty.
The use of MNR as a baseline for future
scenario development does not reflect any
presumption, in favor of MNR or any other
remedy.
From Validation to Prediction -
Monitored Natural Recovery
Forecasts
Contaminated sediment model simulations
begin with initial conditions of sediment
concentrations, add external forcings
including flows and temperatures, and
generate time series of predicted
concentrations for media of concern. During
model development this is done for
calibration and validation periods, which
often coincide with the period of the
remedial investigation (Figure 3).
For remedial planning, forecasting is
extended into the future. The calibration or
validation runs can be continued beyond the
remedial investigation period into the future,
or the simulation can be restarted if current
sediment concentrations are available to reset
initial conditions. Future flows and other
weather-related forcings are unknown, but
can generally be expected to follow patterns
similar to past data unless major system
changes have occurred (e.g., changes in dam
operations, watershed development). To
simulate future years, past time series can be
recycled through the forecast period.
The result of the MNR simulation is a
forecast of surface sediment, water column,
and tissue concentrations. This forecast is
our best estimate of potential future
exposures, under an MNR remedy. To
translate these predicted exposures into time
series of expected human health and
ecological risks, risk factors developed in the
baseline risk assessment can be applied. The
objective of protection of human health and
the environment amounts to reducing this
series of current and future risks.
The Role of Solids in Forecasts
Solids play a key role in chemical fate and
transport at contaminated sediment sites, and
physical and chemical processes involving
solids have been emphasized throughout the
discussion above. We revisit those processes
to discuss a critical element of forecasting,
namely how models simulate the adsorbed
concentrations of chemicals on solids.
Regardless of the remedy selected, solids
accumulation is likely to occur in portions of
the remediated area, and the concentrations
of contaminant on those solids will be a
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
27
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
major driver of post-remedial exposure and
risk. Indeed, with both MNR and more
active remedies, the simulated time trend of
those concentrations can determine the
expected time to achieve remedial goals.
Contaminant concentrations on fresh solids
deposits are determined as follows. Some of
the solids that move down a river in
suspension are present because of upstream
erosion. Solids are deposited at lower flows
and eroded at higher flows, and net
accumulation or depletion of solids depends
on which flows predominate. Models
simulate a depth of erosion as a function of
flow, flow history, and sediment physical
properties (i.e. is sediment fresh and fluffy or
older and more compacted), and ideally do
so using data (where sediments are cohesive)
from flume studies using site-specific
sediment samples (p. 19). These solids are
then tracked by models as suspended in the
water column or settled at downstream
locations, as stream velocities decline with
time or at a widening of the channel cross-
section.
Along with eroded solids, additional solids
enter waterbodies from upstream, from
tributaries and shoreline erosion. Those loads
can be strongly related to flow, and the
relationship between flow and tributary loads
can be determined by monitoring or by
watershed modeling, using land-use and
rainfall data. Solids loads from smaller
tributaries that are not monitored or modeled
can be estimated as proportional to similar
subwatersheds, scaled according to ratios of
the subwatershed areas. It should be
emphasized that accounting for watershed
solids is critical to simulating long-term
system behavior, because long-term changes
in the sediment surface depend on solids in
minus solids out of the system.
Suspended solids are a mix of contaminated
material from eroded deposits and watershed
solids, which are typically less contaminated
(one notable exception being the potential
input of contaminated solids from storm
drain outfalls). Models simulate the mixing
of those solids to produce an average
concentration of contaminant on solid
particles. At the same time, models simulate
some contaminant desorption from solids
into surface water, or adsorption in the
opposite direction, according to an
equilibrium partitioning relationship that
typically leaves most of the contaminant
associated with the solids. Models then
simulate settling such that it is greatest at the
lowest stream velocities.
This is how models simulate concentrations
and deposition rates of freshly deposited
sediment. If erosion is primarily from the top
layer of sediments, then the newly deposited
surface sediments will have some
contamination, but possibly at reduced levels
because of dilution by cleaner watershed
solids. Under extreme conditions, if
sediments are eroded deeply into buried
legacy contamination, surface concentrations
could be increased, in spite of dilution by
watershed solids.
Forecasts of Alternative Remedies
To assess the long-term risk reduction
benefits of alternative remedies, parallel
forecasts can be generated and translated into
alternative risk forecasts. These are similar to
the MNR forecast, using the same forcing
functions and calibrated model parameters,
but with updated bed conditions
incorporating the expected effect of the
remedy on baseline conditions. The long-
term natural processes of burial, mixing, and
dilution by watershed sediments can enhance
the effectiveness of active remedies, and
model simulations forecast the extent to
which that would be expected to occur.
How to Specify Post-remedial Conditions
One critical aspect of the active remedial
simulations in a Feasibility Study is that
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
28
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
updated bed conditions must be assumed
rather than measured, because the remedy
has not yet taken place (and most of the
simulated alternatives will not be
implemented.) This adds an additional layer
of uncertainty that is not present in the MNR
forecast. These starting conditions for
remedial simulations are not generated by
the contaminant transport model; they are
computed offline based on knowledge of the
remediation technology and site conditions.
In addition, the updated bed characteristics
that are specified may be outside of the range
of conditions to which the model has been
calibrated and validated.
To minimize this additional uncertainty,
assumptions about modified bed conditions
should use the best available quantitative
tools. Available tools are discussed below. In
general, the performance of the remedy
should be modeled based on a thorough
synthesis of pilot performance at the site and
the performance of remedies at other sites
with similar conditions.
For example, engineered caps and post-
dredging residual covers can be expected to
change site conditions in these ways:
reduced surface sediment contaminant
concentration
coarser surface (sand or gravel)
reduced organic carbon content, and
reduced water depth.
The coarseness of the surface material is
usually by design, to resist erosion and
increase physical stability of the sediment
bed. Modeling of the shear stresses to which
caps will be subjected under high flows is
commonly done as part of remedial design.
Modeling of short- and long-term risk
reduction due to capping also helps in
evaluating the potential benefits of this
option.
Sands and gravels have higher permeabilities
and lower carbon content than native silts
and clays, increasing the potential for
movement of contaminated pore water into
clean cap materials. Contaminant movement,
if likely to present a problem, can be
addressed in several ways during cap design,
including by adding materials with greater
organic carbon or other sorbent to the cap.
The modeling of pore water transport
through the cap can be performed using a
satellite model specifically designed for this
purpose (Palermo et al, 1998, Reible and
Marquette, 2009).
Unless placed in combination with dredging,
caps also reduce water depth, thereby
reducing the channel cross-section. By
incorporating these bathymetric changes in a
hydrodynamic model, any resulting increases
in stream velocity and/or flooding can be
predicted.
The following conditions can be expected to
result from environmental dredging:
removal of sediment to a target depth
some generated residuals left in place
resuspension of sediment to the water
column, containing adsorbed
contaminant
release of dissolved contaminant to the
water column via desorption from
sediment or pore water release
some undredged inventory left in place,
and
a sand cover or backfill, when specified
as part of the design.
The extent to which generated residuals are
left in place will depend on the
concentrations of the material being dredged,
the difficulty of removing it cleanly (as may
be affected by debris or underlying bedrock),
and the limits of the dredging technology
(NRC 2007b, Bridges et al., 2008). Some
undredged inventory may also be left in
place, depending on the completeness of
sediment characterization and the tolerance
of the post-construction monitoring program.
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
29
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
The expected extent and vertical
distributions of these post-dredging
inventories, based on knowledge of pre-
remedial deposits and experience at sites
with similar conditions and dredging
techniques, should be estimated and used as
updated bed conditions. A practical approach
to estimating residual concentrations a priori,
based on experience at other sites, is outlined
in Palermo et al. (2008).
This is important information for sound
remedial planning, whether or not models
are used to support that planning. The more
realistic and better informed the assumed bed
conditions in remedial forecasts, the more
likely that the long-term consequences of the
remedy can be reliably predicted.
Similarly, updated bed conditions must be
assigned when sand covers, engineered caps,
reactive caps, or other remedial technologies
are simulated. Formulating updated bed
conditions for sand cover and capping
alternatives is more straightforward than for
dredging, because the resistance of sands,
gravels, and stone of given diameters is well
known, as are their contaminant
concentrations. Nevertheless, their placement
can cause mixing with soft sediment, and
ideal placement should not be assumed
without justification. Pilot studies of
remedial technologies can provide very
valuable information for setting updated bed
conditions, thereby reducing uncertainty in
forecasts under remedial alternatives.
Some release of contaminant to the water
column is also expected during remediation,
and the resulting short-term exposures and
contaminant export can have an effect on
bioaccumulation and risk. Estimates of
contaminant loss during dredging range from
less than 1% to up to 9% of contaminant
mass dredged, and will depend on sediment
properties, vertical contaminant distribution,
dredging methods, and salinity (EPA 2005).
Releases to the water column can be
evaluated through a pilot dredging study or
from experience at sites with similar
characteristics and remedial technologies.
How to Model Extreme Events and Evaluate
Long-term Effectiveness
Long-term effectiveness is an essential
component of any contaminated sediment
remedy. This is assessed by evaluating each
remedy under conditions of high erosive
shear stress, which could occur due to an
extreme flow event, wind waves, or tidal
surges (which have a close counterpart in
fresh water bodies in the form of seiches.)
These extreme events should be considered
in the Feasibility Study evaluation of long-
term effectiveness and permanence of the
remedy as well as in the remedial design. It
is likely that the most extreme events
measured in the calibration and validation
data sets fall short of the extremes expected
over the lifetime of the remedy. It is
common to consider the effects of a flow
event expected to occur once every 100
years, on average, and other extreme events
having similar recurrence intervals.
Such an event can be inserted into a long-
term forecast to assess its effect on exposure
and risk. The hydrodynamic model simulates
local stream velocities for the event and
resulting shear stresses on the bed. Where
native material is left in place, as in MNR,
the site-specific credibility relationship that
is already built into the model can be used to
predict the depth of erosion and resulting
release of contaminant. Where specific sizes
of sand or gravel are placed on the bed for
capping or for a post-dredging residual
cover, the modeler can consult the widely
published "Shields Curve" to obtain the
shear stress needed to cause erosion.
The consequences of an extreme event are
likely to be greater for post-MNR than for
post-dredging, because of smaller
inventories of buried contaminant in most
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
30
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
post-dredging scenarios. Simulated effects of
extreme events on caps and covers depend
primarily on their thickness, selected grain
sizes, and erosive forces. Simulations of their
permanence under extreme events can be an
aid to conceptual design, supporting
estimates of the cost to provide sufficient
protectiveness.
Comparisons of Risk Reduction Under
Competing Alternatives
With parallel simulations of possible
remedies in place, the predicted outcomes
can be compared. These take the form of
time lines of exposures and resulting risks. In
making this comparison, differences in time
to plan and implement the different remedial
alternatives should be taken into account,
because these project phases can take many
years for active remedies at large sites. When
compared to MNR, some remedies may
show long-term gains offsetting short-term
setbacks. This can be the case, for example,
where the alternative involves dredging high
concentrations of buried sediment using
technologies that release contaminants and
leave some generated residuals in place,
thereby at least temporarily increasing
surface sediment and fish tissue contaminant
concentrations. In such cases, the assessment
of expected future risk should take both the
short- and long-term effects into account,
weighing a temporary increase in risk against
a long-term reduction. At large sites, it may
also be useful to model the short-term and
long-term effects of combination remedies
by varying the proportions and locations of
MNR, capping, and dredging.
A comparison of the simulations of the
various alternatives in the context of the
National Contingency Plan criteria can
support the selection of a remedy for the site.
Comparisons for the various remedies of
predicted contaminant concentrations in
surficial sediments, the water column, and
fish tissue represent expected differences in
future contaminant exposure and risk. As
suggested in the discussion of uncertainty
above, absolute concentrations and risk
levels predicted for the alternatives are
subject to significant uncertainty on time
scales of decades. Nevertheless, a
comparison of model-predicted trends in risk
under the available alternatives provides a
useful means of quantifying overall risk
reduction. Using process-based numerical
models to do this ensures that decision-
making is consistent with our best
understanding of site data and long-term
processes acting to increase or decrease
exposures over time.
Simulations of the selected remedy are also a
valuable asset when it comes time to monitor
the effectiveness of the remedy. Model
simulations set expectations for post-
remedial time trends in concentrations and
body burdens, as they are expected to vary
from one location to another across the site.
In this way model simulations can help in
designing a monitoring plan, and also in
informing the 5-year reviews of remedy
protectiveness. Where recovery is faster or
slower than expected, models can be
recalibrated to the monitoring data, and if
warranted, used to amend the selected
remedy. From the initiation of the
investigation to the completion of the
remedy, process-based numerical models can
help to test and refine our understanding of
site contamination and the development of a
cost-effective solution that achieves long-
term protection while minimizing short-term
adverse impacts.
Contact Information
For questions on this fact sheet,
please contact Stephen Ells of
OSRTI at Ells.steve@epa.gov or
(703) 603-8822
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
31
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
Using Models to Predict Remedy Effectiveness
A key function of a well-developed model is to aid in assessing the effectiveness of different
remedy alternatives. Rather than choosing between a single major remedy implementation
method (dredging, capping, MNR), models are often used to support evaluation of different
combinations of remedies. At this point in the FS process, models provide input into economics
of remedial alternatives, helping to find the best combination of remedy effectiveness and cost
control. The example below describes how a model can be used to evaluate atypical
combination remedy.
Example:
Background: At a riverine discharge into a coastal zone, historical data on PAH contamination
show generally recovering sediments with some zones of persistent elevated contaminant
concentration. Because of the effects of progressive urbanization of the watershed and
changing hydrology, some of these areas have been identified as vulnerable to erosion under
future extreme events.
Model Development: Based on the historical record of PAH data, academic studies of
sediment loads to the river, and several phases of Rl investigation, an EFDC model is used to
document recovery of surficial sediments. The model is well calibrated, based on reach-based
estimates of deposition and burial, plus site-specific bioturbation rate measurements that
constrain surficial mixing and dilution. The model is developed to a resolution sufficient to
capture remedy implementation at the 100-foot scale. Building on the exposure fields generated
with the EFDC model results, a Gobas model is developed and calibrated to fish body burdens.
Remedy Representation: The proposed remedies include combinations of MNR, capping, and
localized "hot spot" dredging. A key role of the model is to identify the relative benefits of
different combinations of the remedial alternatives.
Model implementation procedures.
MNR: The process of model development and testing build strong stakeholder acceptance of
the model's ability to represent the bioturbation, deposition and burial processes that affect the
viability of MNR. Incorporation of changing hydrology due to watershed development indicates
that recovery processes are expected to continue well into the future, but also directs attention
to a few areas of elevated contaminant levels and potential for exposure under extreme event
conditions.
Capping; Areas of potential vulnerability to future erosion are addressed with a capping
remedy. The remedy is represented in the model by zeroing out resuspension in affected model
cells. Transport through the cap is limited to slow diffusion through the cap, which is simulated
with a simplified offline submodel. Based on the findings of the submodel, diffusion through the
cap is judged to be insignificant and is zeroed out in the EFDC model.
Dredging; In a few areas of significantly elevated concentrations, dredging is recommended to
meet stakeholder objectives for risk management. The remedy is implemented in the model by
assuming a clean sand backfill (background concentrations).
Habitat creation; The remedy also addresses the impact of habitat limitations on the fish
population. Habitat creation is focused on areas of high restoration potential that coincide with
very low contaminant levels. In the Gobas model, habitat preference factors for fish are
modified to favor newly created habitat areas in remediated areas.
United States Office of Superfund Remediation and OSWER Directive 9200.1-96FS
Environmental Protection Technology Innovation November 2009
Agency
32
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
REFERENCES
Bierman, V.J., J.V. DePinto, T.C. Young, P.W. Rodgers, S.C. Martin, R. Raghunathan, S. Hinz,
1992. Development and Validation of an Integrated Exposure Model for Toxic Chemicals in
Green Bay, Lake Michigan. Final Report, Cooperative Agreement CR-814885 U.S
Environmental Protection Agency. September, 1992.
Bridges,T.S., S. Ells, D. Hayes, D. Mount, S.C. Nadeau, M.R. Palermo, C. Patmont, and P.
Schroeder. 2008. The Four Rs of Environmental Dredging: Resuspension, Release, Residual, and
Risk. US Army Corps of Engineers ERDC/EL TR-08-4, February.
Butcher, J.B., E.A. Garvey, VJ. Bierman. 1997. Equilibrium Partitioning of PCB Congeners in
the Water Column: Field Measurements from the Hudson River. Chemosphere: 36 (15).
D. Glaser and T. Bridges. 2007. Separating the Wheat from the Chaff: the Effective Use of
Mathematical Models as Decision Tools. Integrated Environmental Assessment and
Management 3:442-449
Mackay, D. et al. 2006. Handbook of Physical-Chemical Properties and Environmental Fate for
Organic Chemicals, Second Edition. CRC Press, Boca Raton.
NRC. 2007a. Models in Environmental Regulatory Decision Making. Committee on Models in
the Regulatory Decision Process, National Research Council. 286 pp.
NRC, 2007b. Dredging at Superfund Megasites: Assessing the Effectiveness. National Research
Council of the National Academies, Committee on Sediment Dredging at Superfund Megasites,
National Research Council, 236 pp.
Palermo, M., Maynord, S., Miller, L, and Reible, D. 1998. "Guidance for In-Situ Subaqueous
Capping of Contaminated Sediments," EPA 905-B96-004. Great Lakes National Program Office,
Chicago, IL.
Palermo, M., Schroeder, P.R., Estes T.J., Francingues, N.R. 2008. Technical Guidelines for
Environmental Dredging of Contaminated Sediments, US Army Corps of Engineers. ERDC/EL
TR-08-29, September.
Reible, D. and A. Marquette. 2009. Capping Design Model. Southwest Hazardous Substance
Research Center, http://capping.ce.utexas.edu/design.html
USEPA. 2005. Contaminated Sediment Remediation Guidance for Hazardous Waste Sites.
Office of Solid Waste and Emergency Response. EPA-540-R-05-012. OSWER 9355.0-85.
December.
USEPA. 2008a. Guidance on the Development, Evaluation and Application of Environmental
Models. Council for Regulatory Environmental Modeling, U.S. Environmental Protection
Agency, Washington, DC 20460, 89 pp .
USEPA, 2008b. Using Fish Tissue Data to Assess Remedy Effectiveness: Sediment Assessment
and Monitoring Sheet (SAMS) #1. Prepared under OSWER Directive 9200.1-77D, July, 2008
USGS, 1998. Ground Water and Surface Water, a Single Resource. USGS Circular 1139.
Denver, CO.
United States Office of Superfund Remediation and OSWER Directive 9200.1-96FS
Environmental Protection Technology Innovation November 2009
Agency
33
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
APPENDIX A:
A Partial Listing of Available Hydrodynamic, Sediment Transport,
Contaminant Transport, and Food Chain / Ecological Effects Models
Hydraulic/Hydrodynamic/Sediment Transport/Contaminant Transport Models
CE-QUAL-W USAGE Dynamic, 2D Laterally-Averaged Water Quality Model
CH3D-SED USAGE Dynamic, 3-D Curvilinear Hydrodynamics and Sediment Transport
Model
CORMIX USEPA 3-D Steady-State Analytical Mixing-Zone Model
DYNHYD5 USEPA/CEAM Dynamic 1-D Link-Node Tidal Hydrodynamic Model
ECOMSED HQI Dynamic, 3-D Hydrodynamic and Sediment Transport Model
EFDC Tt/VIMS/EPA 3-D Environmental Fluid Dynamics Code
HEC-RAS USAGE Dynamic, 1-D River Analysis System
HEM1-3D VIMS Dynamic, 1-3D Hydrodynamic and Eutrophication Models
HSCTM-2D USEPA/CEAM 2-D Dynamic Hydrodynamic Sediment and Contaminant
Transport Model
RCA HQI Dynamic Water Quality Simulation Model
RIVMOD-H USEPA/CEAM Dynamic, 3-D River Hydrodynamic Model
RMA-2V WES Dynamic, 2-D Hydrodynamic Analysis Model
WASP5/6/7 USEPA Dynamic Water Quality Simulation Model
Hydrologic/Watershed Models
AGNPS
BASINS
HSPF
SWAT
SWMM
WARMF
USDA Agricultural Non-Point Source Pollution Model
USEPA/CEAM Point and Non-Point Source Model Toolbox
USEPA/CEAM Simulation of Mixed Land-use Watersheds
USDA Soil and Water Assessment Tool
USEPA Stormwater Management Model
EPRI Watershed Analysis Risk Management Framework
Food Chain / Ecological Effects Models
AQUATOX USEPA Ecosystem / Food Web Bioaccumulation Model
AQUAWEB Arnot and Gobas Food Web Bioaccumulation Model
BASS Bioaccumulation and Aquatic System Simulator
ECOFATE Gobas (1993) Model of Ecological Fate and Food Web Bioaccumulation
EXAMS II USEPA/CEAM Fate and Exposure Model for Assessing Toxics
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
34
OSWER Directive 9200.1-96FS
November 2009
-------
Understanding the Use of Models in Predicting the Effectiveness of Proposed Remedial Actions at
Superfund Sediment Sites
FGETS USEPA/CEAM Food and Gill Exchange of Toxic Substances Fish
Bioaccumulation Model
HEP/HS USEPA/CEAM Habitat Evaluation Procedures/Habitat Suitability Indices
HES USEPA/CEAM Habitat Evaluation System Used to Assess the Impacts of
Development Projects for Aquatic and Terrestrial Habitat Evaluations
IFIM USEPA/CEAM Instream Flow Incremental Methodology for Riverine
Habitats
PHABSIM USEPA/CEAM Fish-habitat Preference and Discharge-Habitat Model
PVA USEPA/CEAM Population Viability Analyses
QEAFDCHN QEA Food Chain Bioaccumulation Model
RBPs USEPA/CEAM Rapid Bioassessment Protocols
SERAFM Spreadsheet-Based Ecological Risk Assessment for the Fate of Mercury
THOMANN Thomann Fish / Food Web Model
United States
Environmental Protection
Agency
Office of Superfund Remediation and
Technology Innovation
35
OSWER Directive 9200.1-96FS
November 2009
------- |