
ENVIRONMENTAL data
aren’t easy to obtain, and once obtained, they are often hard
to interpret. For example, drilling into the earth to determine
what kind of soil exists at any given spot in the substrate is not
only expensive but also gives scientists just piecemeal information.
Computer analysis with this information can be equally piecemeal.
But earth scientists are learning that computer models can be made
more meaningful when they are stochastic, meaning that they are
based on a certain amount of probability. Now, with the capability
of highperformance supercomputers in the National Nuclear Security
Administration’s Advanced Simulation and Computing (ASCI) program,
Livermore scientists are exploring groundbreaking ideas in statistical
theory that will help them use stochastic descriptions quantitatively
and obtain a much more complete picture of soil composition.
This
new technology, called a stochastic engine, is a process that links
predictive models, advanced statistical methods, and refined search
methods. Using this technology, scientists can incorporate a proposed
soil configuration into a computer model and produce a geophysical
simulation. The simulated result is compared to actual data. If
the result is consistent with observed data, then the simulation
is boosted to the next phase of analysis.
The stochastic method is
a powerful technique that is now in use. Livermore scientists are
consulting on a project with the Westinghouse Savannah River Company
in which the stochastic engine will assist in a major cleanup operation
at the Savannah River Site in South Carolina. The method could also
be applied to problems in stockpile stewardship, atmospheric dispersion,
seismic velocities, and intelligence collection.
Cleanup Site Yields New Tool
The stochastic engine
concept uses techniques developed at Livermore and was motivated
by an innovative steam remediation cleanup being conducted by Southern
California Edison at a Superfund site in Visalia, California, in
which Laboratory scientists also participated. (See S&TR,
January/ February 1996,
Groundwater
Cleanup Using Hydrostratigraphic Analysis.) During the project,
more than 46 million pieces of data were obtained pertaining to
the way steam, water, and contaminant flowed through the groundwater
plumbing system. These data included temperatures, flow rates, pressures,
and electrical resistance tomography (ERT) measurements. ERT, a
technology developed at Livermore in 1993 and now available commercially,
is similar to a computed tomography scan. It images soil resistivity,
and that gives scientists information on soil properties such as
temperature, soil type, and saturation. While the data collected
from Visalia were rich and invaluable for Edison’s operational
decisions, the various data types could only be used independently.
Observations and simulations could not be linked to provide the
kind of cohesive understanding that would dramatically improve site
operations and, most importantly, optimize the final outcome of
the cleanup work.
The work at Visalia, while
highly successful overall, is representative of a frustration that
Livermore environmental scientists experience whenever they attempt
to characterize soil compositions at cleanup sites: how to apply
the powerful predictive capabilities of Livermore’s supercomputers
to complex, real situations. For the past year, Roger Aines and
a multidisciplinary team have been discussing how to apply modern
computational power and statistical search methods to extract maximum
information from sparse initial data and then to improve the analysis
on the fly as more data become available.

(a) Images of the soil at
the Savannah River Site obtained by conventional electrical
resistance tomography (ERT) show electrical properties of the
site based on data that have been smoothed (its differences
have been evened out), and (b) a stochastic engine analysis
of the same data, which shows the probability of the three local
soil types (sand, silt–clayey sand, and clay) at each location. 
More Than One Right Answer
The power of the stochastic engine comes from
its ability to refine a model by successively narrowing down the
possible configurations of a hypothetical model. The refinement
is done over progressive layers of data. In this process of model
improvement through iteration, the stochastic engine uses an advanced
statistical method called a hybrid Markov Chain Monte Carlo (MCMC)–Bayesian
analysis.
In the MCMC analysis, a
chain (or sequence) of configurations is considered. Each configuration
undergoes
a probability calculation that compares observed data to corresponding
model predictions. If the predictions are acceptable (that is, probable
for the configuration), the result of that calculation becomes the
basis of the next configuration. This allows the process to rapidly
search for good configurations in very complex situations. The Bayesian
statistical method, based on the work of English mathematician Thomas
Bayes, performs its part in the stochastic engine by comparing the
probability calculations with real information to guide the statistical
inference process.
Suppose a volume of soil
is known to be composed of seven layers that could be either sand
or silt, and an ERT measurement of that volume gives a value of
11. The stochastic approach calculates which configurations of silt
and sand, and in which positions, give values close to 11. Each
case with a value near 11 is passed on to the next stage of analysis.
There, the model will continue to restrict possible configurations
but base its decisions on other data types, such as water, temperature,
or pressure.
For the simple case cited
here, it is easy to calculate and compare all the possible configurations,
but for a large area, such as the Visalia cleanup site, the possibilities
are far too numerous. At Visalia, the MCMC–Bayesian method
could help by performing an efficient intelligent search through
the collection of possible soil configurations, rapidly identifying
the configurations that most closely match all the data.
“It’s not about
trying to find the single best answer, but all of the good answers,”
says Aines. “In underground problems, there are usually multiple
solutions that are consistent with the data.”
The stochastic engine’s
ability to choose system configurations that are consistent with
observed data allows much more tightly constrained (better restricted)
answers than conventional methods. Only the ways the system can
possibly exist are considered. Using the stochastic technique, for
example, researchers can interpret ERT images to derive characteristic
soil types for a site, rather than simply provide the electrical
properties of the ground. The stochastic engine allows the available
information to be used more effectively. It also allows the user
to incorporate known constraints, such as the presence of a gravel
layer observed in a well, to further guide the statistical inference.

The sewer at Savannah River
Site, South Carolina, that has been the source of solvents soaking
into the soil. At right, geoscientist Steve Carle is shown at
the sewer outfall. He was examining soil just below the outfall,
looking at the silt layers that tend to control the migration
of the solvent. 
It Doesn’t Have to End with Dirt
The stochastic engine method
has tremendous potential for use in disciplines that need to combine
data and simulation. Currently, the team is working with a number
of scientists from other Livermore directorates to identify unknown
sources of toxic contaminants in the atmosphere, locate flaws in
buildings, evaluate intelligence data, and expand tomography and
xray imaging data.
The Savannah River Site project
illustrates how the engine is being used in industrial partnerships.
Livermore has been consulting with Westinghouse’s Savannah
River Company to clean up organic solvents from the soils and groundwater
at the South Carolina site.
Since
1983, the company has been performing environmental cleanup of a
site where, over time, solvents became a solvent plume that extended
over 5 square kilometers. Now, Westinghouse is ready to present
its cleanup results to regulators and assure the community that
the remaining plume will not affect surface water bodies. The stochastic
engine will be used to evaluate the effectiveness of source cleanup
and to predict the ultimate effect of the remaining plume.
Challenges
Ahead
Why hasn’t the stochastic
method been used before? For one thing, the complexity of the method
has required robust computer power that simply has not been available
until recently. For another, even with the power available on ASCI
computers, some scientists are still skeptical of the method. Aines
says that because underground problems are so complex, many people
are displaying a “show me first” attitude toward the technology.
“No one has done this before, so some believe it can’t be
done.” The Savannah River Site project may prove that the engine
is a feasible and valuable tool for environmental cleanup and more.
—Laurie Powers
Acknowledgments:
The stochastic engine work described in this article was developed
by a multidisciplinary team from Livermore’s Energy and Environment,
Engineering, and Computation directorates, including computational
geoscientists John Nitao and Steve Carle; engineering statisticians
Bill Hanley, Ron Glaser, and Sailes Sengupta; geophysicists Robin
Newmark and Abe Ramirez; and computational scientist Kathy Dyer.
Key Words: Bayesian
statistics, electrical resistance tomography, Markov chain, Monte
Carlo method, Savannah River Site, stochastic engine, Superfund,
Visalia cleanup.
For further information contact Roger Aines (925) 4237184 (aines1@llnl.gov).
