DATA MINING WITH DIFFERENT TYPES OF X-RAY DATA

High-Throughput Materials Discovery uses automation and parallelism to synthesize and evaluate large numbers of specimens while reducing time and costs associated with finding and optimizing novel materials. As optimal performance may not be uniformly distributed throughout parameter space, efficient tools for analyzing data and evaluating large areas of compositional or parameter space are needed. Data mining tools enable moving from the statistics of limited experimental designs to more descriptive and predictive relationships. Clustering a set of 47 samples for which both X-ray powder diffraction data and X-ray fluorescence-based elemental composition data were available showed that elemental composition correlated strongly with phase composition in this particular set of samples. Also, the clustering of the X-ray data was found to be exactly coincident with a different sample characteristic "type". Decision tree classification of a larger data set of 86 samples showed that "type" could be defined with very few errors from relatively few splits of the XRF-based compositions. Although composition exhibited strong clustering, measures of performance in these same samples exhibited only very weak clustering. However, performance of the materials could be predicted from linear regression using different slices of the data. Neural nets were attempted for improved predictability of performance beyond linear regression. As expected from the liner regression results, single output linear-based multi-layer perceptrons yielded acceptable predictive capability, but were found to yield notably degraded predictive results if "type" was excluded from the models. The strong dependence of performance on "type" for these samples was an unexpected outcome of the data analysis.