AbstractFormal optimization techniques, of the kind developed in operations research to aid business decision‐making, show promise as tools for controlling the high costs of ground‐water remediation. A major barrier to the adoption of these tools in the industry, however, is the burden they place on the computational resources of an organization. Whenever a new pumping strategy is considered, a flow and transport model must be called to evaluate its effectiveness. Since it is common for optimization techniques to examine hundreds of strategies, at the very least, the bottleneck caused by the modeling step quickly becomes prohibitive. A method has been developed which uses artificial neural networks to substitute for the full model, permitting millions of strategies to be evaluated within reasonable time‐frames on conventional computer equipment. This method was applied to a 28‐location pump‐and‐treat problem at a western Superfund site. Networks were first trained to predict mass‐extraction and containment information normally generated by the 2‐D model SUTRA. They were then used in an optimization procedure to identify 250 (out of over 4 million) pumping patterns which met restoration goals at minimum cost. Analyses of the winning patterns classified individual locations as popular vs. unpopular. Sensitivity analyses on each location further distinguished: (1) locations which were strong performers under most circumstances from those which depended on other locations, and (2) locations which didnt aid the cleanup from those which were actually detrimental to it. All these analyses were made possible by the rapid evaluation of patterns provided by the
展开▼
