Optimal experimental design for materials discovery

Abstract

In this paper, we propose a general experimental design framework for optimally guiding new experiments or simulations in search of new materials with desired properties. The method uses the knowledge of previously completed experiments or simulations to recommend the next experiment which can effectively reduce the pertinent model uncertainty affecting the materials properties. To illustrate the utility of the proposed framework, we focus on a computational problem that utilizes time-dependent Ginzburg-Landau (TDGL) theory for shape memory alloys to calculate the stress-strain profiles for a particular dopant at a given concentration. Our objective is to design materials with the lowest energy dissipation at a specific temperature. The aim of experimental design is to suggest the best dopant and its concentration for the next TDGL simulation. Our experimental design utilizes the mean objective cost of uncertainty (MOCU), which is an objective-based uncertainty quantification scheme that measures uncertainty based upon the increased operational cost it induces. We analyze the performance of the proposed method and compare it with other experimental design approaches, namely random selection and pure exploitation.