# The Stochastic Engine Initiative: Improving Prediction of Behavior in Geologic Environments We Cannot Directly Observe

## Abstract

The stochastic engine uses modern computational capabilities to combine simulations with observations. We integrate the general knowledge represented by models with specific knowledge represented by data, using Bayesian inferencing and a highly efficient staged Metropolis-type search algorithm. From this, we obtain a probability distribution characterizing the likely configurations of the system consistent with existing data. The primary use will be optimizing knowledge about the configuration of a system for which sufficient direct observations cannot be made. Programmatic applications include underground systems ranging from environmental contamination to military bunkers, optimization of complex nonlinear systems, and timely decision-making for complex, hostile environments such as battlefields or the detection of secret facilities. We create a stochastic ''base representation'' of system configurations (states) from which the values of measurable parameters can be calculated using forward simulators. Comparison of these predictions to actual measurements drives embedded Bayesian inferencing, updating the distributions of states in the base representation using the Metropolis method. Unlike inversion methods that generate a single bestcase deterministic solution, this method produces all the likely solutions, weighted by their likelihoods. This flexible method is best applied to highly non-linear, multi-dimensional problems. Staging of the Metropolis searches permits us to run the simplest modelmore »

- Authors:

- Publication Date:

- Research Org.:
- Lawrence Livermore National Lab., CA (US)

- Sponsoring Org.:
- US Department of Energy (US)

- OSTI Identifier:
- 15002143

- Report Number(s):
- UCRL-ID-148221

TRN: US200408%%139

- DOE Contract Number:
- W-7405-ENG-48

- Resource Type:
- Technical Report

- Resource Relation:
- Other Information: PBD: 9 May 2002

- Country of Publication:
- United States

- Language:
- English

- Subject:
- 54 ENVIRONMENTAL SCIENCES; ALGORITHMS; CONFIGURATION; CONTAMINATION; DECISION MAKING; EFFICIENCY; ELECTRIC CONDUCTIVITY; ENGINES; FORECASTING; GEOCHEMISTRY; LITHOLOGY; NONLINEAR PROBLEMS; OPTIMIZATION; PLUMES; SUPERCOMPUTERS; TOMOGRAPHY

### Citation Formats

```
Aines, R, Nitao, J, Newmark, R, Carle, S, Ramirez, A, Harris, D, Johnson, J, Johnson, V, Ermak, D, Sugiyama, G, Hanley, W, Sengupta, S, Daily, W, Glaser, R, Dyer, K, Fogg, G, Zhang, Y, Yu, Z, and Levine, R.
```*The Stochastic Engine Initiative: Improving Prediction of Behavior in Geologic Environments We Cannot Directly Observe*. United States: N. p., 2002.
Web. doi:10.2172/15002143.

```
Aines, R, Nitao, J, Newmark, R, Carle, S, Ramirez, A, Harris, D, Johnson, J, Johnson, V, Ermak, D, Sugiyama, G, Hanley, W, Sengupta, S, Daily, W, Glaser, R, Dyer, K, Fogg, G, Zhang, Y, Yu, Z, & Levine, R.
```*The Stochastic Engine Initiative: Improving Prediction of Behavior in Geologic Environments We Cannot Directly Observe*. United States. doi:10.2172/15002143.

```
Aines, R, Nitao, J, Newmark, R, Carle, S, Ramirez, A, Harris, D, Johnson, J, Johnson, V, Ermak, D, Sugiyama, G, Hanley, W, Sengupta, S, Daily, W, Glaser, R, Dyer, K, Fogg, G, Zhang, Y, Yu, Z, and Levine, R. Thu .
"The Stochastic Engine Initiative: Improving Prediction of Behavior in Geologic Environments We Cannot Directly Observe". United States. doi:10.2172/15002143. https://www.osti.gov/servlets/purl/15002143.
```

```
@article{osti_15002143,
```

title = {The Stochastic Engine Initiative: Improving Prediction of Behavior in Geologic Environments We Cannot Directly Observe},

author = {Aines, R and Nitao, J and Newmark, R and Carle, S and Ramirez, A and Harris, D and Johnson, J and Johnson, V and Ermak, D and Sugiyama, G and Hanley, W and Sengupta, S and Daily, W and Glaser, R and Dyer, K and Fogg, G and Zhang, Y and Yu, Z and Levine, R},

abstractNote = {The stochastic engine uses modern computational capabilities to combine simulations with observations. We integrate the general knowledge represented by models with specific knowledge represented by data, using Bayesian inferencing and a highly efficient staged Metropolis-type search algorithm. From this, we obtain a probability distribution characterizing the likely configurations of the system consistent with existing data. The primary use will be optimizing knowledge about the configuration of a system for which sufficient direct observations cannot be made. Programmatic applications include underground systems ranging from environmental contamination to military bunkers, optimization of complex nonlinear systems, and timely decision-making for complex, hostile environments such as battlefields or the detection of secret facilities. We create a stochastic ''base representation'' of system configurations (states) from which the values of measurable parameters can be calculated using forward simulators. Comparison of these predictions to actual measurements drives embedded Bayesian inferencing, updating the distributions of states in the base representation using the Metropolis method. Unlike inversion methods that generate a single bestcase deterministic solution, this method produces all the likely solutions, weighted by their likelihoods. This flexible method is best applied to highly non-linear, multi-dimensional problems. Staging of the Metropolis searches permits us to run the simplest model systems, such as lithology estimators, at the lower stages. The majority of possible configurations are thus eliminated from further consideration by more complex simulators, such as flow and transport models. Because the method is fully automated, large data sets of a variety of types can be used to refine the system configurations. The most important prerequisites for optimal use of this method are well-characterized forward simulators, realistic base representations, and most importantly an ability to obtain disparate data sets that are directly affected by the system configuration. Our initial earth-sciences application uses models for lithology, flow and transport, geochemistry, and geophysical imaging; the system configuration (base representation) being refined is the rock type at each underground location. In the initial stages of this initiative we demonstrated a two-stage analysis of synthetic Electrical Resistance Tomography (ERT) data and hydraulic flow information (Newmark et al., 2002). We used these results to develop algorithms that improve efficiency of the Metropolis search and provide accurate diagnostic evaluation during the search. Using actual data from a highly contaminated A/M outfall and solvent tank storage areas at the Savannah River Site (SRS), we used the stochastic engine to resolve lithology using ERT data. SRS will use these methods in their design and implementation of steam cleanup of the largest trichloroethylene (TCE) source in the Department of Energy (DOE) complex. We have implemented ''soft conditioning'' algorithms that allow us to use a variety of data types to control the initial representations, and most importantly, to use the final distribution resulting from one stochastic engine analysis as the initial distribution for a subsequent analysis. We have created a web-based interface that will allow collaborators like SRS to enter data and observe results of calculations on Lawrence Livermore National Laboratory (LLNL) supercomputers in an interactive mode. All engine functions operate in three dimensions, and a parallel implementation on Linux cluster machines is in initial testing. The method will be extended to include active process analysis, in which an ongoing data stream is used to continuously update the understanding of the system configuration. Applications to other types of state spaces, such as chemical parameters in a reacting system or atmospheric plume movement, are being evaluated.},

doi = {10.2172/15002143},

journal = {},

number = ,

volume = ,

place = {United States},

year = {2002},

month = {5}

}