The SERDP Munitions Response (MR) program area is focused on developing innovative methods to characterize, remediate, and sustainably administer sites contaminated with unexploded ordnance (UXO) or other munitions of concern. The possibility of buried UXO presents a significant challenge to site management, as they are difficult to detect, monitor, and remediate. Of particular concern is the potential for UXO to migrate from their current location into an area with high likelihood of human contact. Due to the complexity of the environment and UXO response, decision making under uncertainty is required.

The objective of this effort was the construction of a predictive Underwater Munitions Expert System (UnMES) providing computer-based decision support for management of aquatic sites.

UnMES is based on a Bayesian Network (BN) framework and acts to synthesize results from multiple research projects supported by the SERDP MR program investigating the behavior of underwater munitions on the seabed.

A BN is a graphical probability model that reflects current knowledge of the probable UXO migration and burial in an underwater environment. The difficulty of obtaining environmental data, unknown details of the physical processes, and the exact disposition of the munitions result in a high level of uncertainty for UXO location and subsequent behavior. A BN is a useful method of modeling systems with complex relationships in a probabilistic manner. A BN captures the uncertainty in relationships between the variables, quantified as the spread in the probability distributions. Predictions can be made with uncertainty in the inputs, and the uncertainty is propagated by probabilistic inference using Bayes theorem.

This requires a large number of data cases. However, due to the limited amount of field and laboratory observations applicable to UXO burial and mobility, an alternative approach was taken to construct UnMES. Deterministic models were developed to capture the first-order physics of the processes of interest, and then utilized to produce datasets of cases linking input conditions to output UXO behavior.

To establish user acceptance in the expert system, evaluation at high confidence levels, such as commonly cited for unidimensional statistics, e.g., 95% Confidence Interval (C.I.), would be desirable. However, the sampling demands for statistics of multidimensional probabilities make those confidence levels difficult to attain. Also, considering the remaining acknowledged uncertainties in the underlying physical models, selection of more modest confidence criteria, such as 90% or even 80% C.I., are suitable for UnMES assessment. Particular effort may be needed to build the user community comfort with the operational value furnished by these more modest validation scores and the use of relative skill metrics.

The development of UnMES for the prediction of underwater munitions burial and mobility has provided a focal point for the collection and synthesis of information obtained by the on-going research projects within the SERDP MR program. Efforts at Johns Hopkins University Applied Physics Laboratory have included laboratory experiments, refinement of physics-based process models, and analysis of multiple field tests in collaboration with colleagues among the SERDP researchers.