Prediction of lithology types at the Hanford 300 Area using a clustering analysis

The purpose of this study is to find an optimal method for mapping the three-dimensional distribution of lithology at the Hanford IFRC site 300 Area based on surrogate measurements. We considered 6 types of measurements for this analysis: gamma ray, concentration of U-238 (609), K-40, U-238 (1764), Th-232, and the hydraulic conductivity. To decide which combinations of variables are best suited for determining lithology type, we trained our classification method using training sets that included several wells with lithological information. A clustering analysis was applied to each training set and the lithology types for each cluster of the training set were fitted with a probability distribution function. The lithology type at each point in the testing set was selected to be the one linked with the mode of the distribution at the corresponding cluster. The predictions were then checked against the data of the testing set. This process was applied repeatedly using different numbers of clusters. In addition, many different configurations of training sets and testing sets were used to establish confidence in the predictive ability of the clustering and classification methods. Our best success rates as measured by matching predictions with observations were obtained for 2 or 3 clusters, and the following measurements: concentration of U-238 (609), K-40, U-238 (1764), and Th-232, and were consistently around 80%.