Evaluating Ground-Water Quality Using Artificial Adaptive Systems

We evaluate applicability of the auto-contractive mapping (AutoCM) technique to identify connectivity among ground-water-quality variables in aquifers worldwide. This bottom-up approach is based on unsupervised training by field parameters, aqueous chemistry, and isotopes sampled from wells in crystalline bedrock, sedimentary, and alluvial aquifers. The neural connectivity is evaluated using a minimum spanning tree and then compared to component planes projections of weight vectors derived from the unsupervised self-organizing map (SOM) technique. The AutoCM spanning tree structure resembles the principal component analysis of the SOM component planes. Branch clusters at extremes of the trees represent strongly correlated variables diminishing father down a branch or among clusters with increasing separation. AutoCM appears promising as a rapid tool for understanding water-quality relations and may possibly be used in the model reduction process, but further testing is required.