Exploring the Generalization Ability of Machine Learning Models: Predicting physical information of materials with unknown elements

Xue Dezhen1

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

EXTENDED ABSTRACT: Design of multicomponent materials encounters a huge combinatorial space. An efficient search in such a space is somehow beyond the capability of the conventional knowledge-guided approaches. Emerging artificial intelligence and machine learning (ML) methods thereby have recently been applied to accelerate the composition design of new materials. Usually the design efficiency relies on a surrogate model given by machine learning (ML), in which a set of materials features, are mapped to a target property. The features, containing information of composition and elements, determine the up-limit of the performance for the surrogate model and are thus crucial to the subsequent design of new materials. In the present study we employ domain knowledge of composition dependence of transformation behaviors in ferroic materials properties to construct new features for the composition design of ferroic materials. The new features effectively improve the performance of machine learning for properties such as phase transition temperature, energy storage density, and piezoelectric response, etc. Furthemore, we propose a feature selection method that maximizes the efficiency of the active learning process, especially where data sizes are limited. The method significantly reduces the number of new experiments to find the desired targets compared with several other feature selection methods. It is robust in terms of selecting features irrespective of the hyper-parameter values, and is applied to several materials datasets including high entropy alloys, shape memory alloys, and ferroelectric ceramics. We synthesize and characterize high entropy alloys with high strength within an active learning loop using features obtained by the proposed method. The results indicate that domain knowledge assisted construction of features together with feature selection using active learning can further improve the efficiency of data-driven design of ferroic materials and high entropy alloys.

Keywords: ferroic materials; high entropy alloys; machine learning; active learning; feature selection 

Brief Introduction of Speaker
Dezhen Xue

Dezhen Xue received the B.Sc. and Ph.D. degrees from Xi'an Jiaotong University (XJTU), Xi'an, China, in 2006 and 2012, respectively. During his Ph.D. thesis, he spent four years at the National Institute of Materials Science, Tsukuba, Japan, for joint research. He was a Directors­Funded Post-Doctoral Fellow at the Los Alamos National Laboratory, Los Alamos, NM, USA, before beginning his independent career at XJTU in 2016. He is currently a full Professor of materials science at XJTU. He has authored more than 100 peer-reviewed papers. His research interests include materials informatics and currently concerns accelerated searching for new materials using machine learning and optimization algorithms.