PNNL

Abstract

With our present concern for a secure environment, the development of new radiation detection materials has focused on the capability of identifying potential radiation sources at increased sensitivity levels. As the initial framework for a materials-informatics approach to radiation detection materials, we have explored the use of both supervised (Support Vector Machines – SVM and Linear Discriminant Analysis – LDA) and unsupervised (Principal Component Analysis – PCA) learning methods for the development of structural signature models. Application of these methods yields complementary results, both of which are necessary to reduce parameter space and variable degeneracy. Using a crystal structure classification test, the use of the nonlinear SVM significantly increases predictive performance, suggesting trade-offs between smaller descriptor spaces and simpler linear models.