Multi-objective Optimization for Materials Discovery via Adaptive Design

Abhijith M Gopakumar,Prasanna V Balachandran,Turab Lookman,Dezhen Xue,James E Gubernatis

doi:10.1038/s41598-018-21936-3

Abhijith M Gopakumar, Prasanna V Balachandran + Show 3 more

Open Access

https://doi.org/10.1038/s41598-018-21936-3

Copy DOI

Abstract

Guiding experiments to find materials with targeted properties is a crucial aspect of materials discovery and design, and typically multiple properties, which often compete, are involved. In the case of two properties, new compounds are sought that will provide improvement to existing data points lying on the Pareto front (PF) in as few experiments or calculations as possible. Here we address this problem by using the concept and methods of optimal learning to determine their suitability and performance on three materials data sets; an experimental data set of over 100 shape memory alloys, a data set of 223 M2AX phases obtained from density functional theory calculations, and a computational data set of 704 piezoelectric compounds. We show that the Maximin and Centroid design strategies, based on value of information criteria, are more efficient in determining points on the PF from the data than random selection, pure exploitation of the surrogate model prediction or pure exploration by maximum uncertainty from the learning model. Although the datasets varied in size and source, the Maximin algorithm showed superior performance across all the data sets, particularly when the accuracy of the machine learning model fits were not high, emphasizing that the design appears to be quite forgiving of relatively poor surrogate models.

Highlights

We demonstrated how machine learning models in conjunction with optimization strategies, can guide the experiments or calculations towards finding materials with desired single objectives or properties[15,16]
The datasets used in this work varied in size, fidelity and source, the Maximin optimization algorithm showed superior performance across these cases in which the accuracy of the machine learning regression model fits were too low to be considered reliable for predictions
The number of design cycles can be restricted by limiting the number of new measurements or when the sub-Pareto front (PF) after a given number of measurements meets the requirements put on the materials properties by the researcher

Summary

Introduction

We demonstrated how machine learning models in conjunction with optimization strategies, can guide the experiments or calculations towards finding materials with desired single objectives or properties[15,16]. Using an adaptive learning paradigm based on active or reinforcement learning ideas from computer science, we showed how to iteratively select or recommend candidates for experiments or calculations and update known training data with each new sample synthesized or computed to subsequently improve the search. New alloys[15] and piezoelectric compositions[16] with desired very low dissipation or phase boundary characteristics were found in this manner. Because of the vast search space and limited training data, the probability of finding these compounds by conventional trial and error approaches is exceedingly low

Objectives

Results

Conclusion