Accelerated search for materials with targeted properties by machine learning and adaptive design

Dezhen Xue^*,1, Xiangdong Ding¹, Jun Sun¹ and Turab Lookman²

1 State Key Laboratory for Mechanical Behavior of Materials, Xi'an Jiaotong University, Xi'an 710049, China

2 Theoretical Division, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA

Abstract

  Finding novel materials in an accelerated, yet cost effective manner that is not dependent on trial and error is one of the central goals of the U.S. Materials Genome Initiative. Learning from data is an attracting tool to accelerate the discovery of new materials. However, in addition to data, a distinguishing aspect of materials science is that there exists a substantial body of knowledge in the form of phenomenological models and physical theories. Here we focused on combining the informatics techniques and materials knowledge to further accelerated search for new materials with targeted properties. Two case studies include 1) using results from Landau–Devonshire theory to guide experiments in the design of new lead-free piezoelectrics with better temperature reliability; 2) using physical understanding to isolate possible global minimums of the search space to achieve a ferroelectric material with higher en

ergy storage density using as few experiments as possible. Our framework may offer the opportunity to significantly reduce the number of costly and time-consuming experiments.

       DOI: 10.12110/firstfmge.20171121.414