(595g) Identifying Descriptors for Materials Science Via Genetic Programming: A Case Study for Dielectric Breakdown Strength

The development of simple phenomenological theories via descriptor identification makes possible rapid screening large numbers of materials and ultimately facilitates the design of new materials. However, determining how to utilize large material data sets to identify the best descriptors for a given property remains one of the biggest challenges in materials science. In this talk, we demonstrate the use of genetic programming to identify descriptors of dielectric breakdown based on 82 representative crystalline materials. Band gap E_g and phonon cut-off frequency w_max were found to be the two most relevant properties, and new classes of phenomenological models featuring functions of E_g and w_max were discovered. The genetic programming model was found to outperform other machine learning models in descriptor identification, and we discussed how this approach may be extended to other problems in materials science.