A mixture of experts based discretization approach for characterizing subsurface contaminant source zones

Accidental releases and improper disposal of hazardous chemicals has led to widespread chemical contamination of subsurface soils and water-bearing formations. Effective remediation and restoration of such contaminated sites is dependent upon knowledge of the contaminant´s mass and distribution within the aquifer. Recent research has shown that the estimation of certain metrics which summarize the distribution of the contaminant in the source-zone is sufficient for designing effective remediation strategies. In this work we explore the task of predicting such a metric based upon down-gradient concentration profiles. Motivated by the underlying physics of this problem we model this as a classification task where each class represents a particular sub-range of the metric. The solution to this problem is obtained by adapting the mixture of experts (MoE) scheme to learn a suitable quantization of the metric. Experimental evidence shows that this scheme outperforms baseline methods.