MDTS: automatic complex materials design using Monte Carlo tree search

Thaer M Dieb,Shenghong Ju,Kazuki Yoshizoe,Zhufeng Hou,Junichiro Shiomi,Koji Tsuda

doi:10.1080/14686996.2017.1344083

Abstract

Complex materials design is often represented as a black-box combinatorial optimization problem. In this paper, we present a novel python library called MDTS (Materials Design using Tree Search). Our algorithm employs a Monte Carlo tree search approach, which has shown exceptional performance in computer Go game. Unlike evolutionary algorithms that require user intervention to set parameters appropriately, MDTS has no tuning parameters and works autonomously in various problems. In comparison to a Bayesian optimization package, our algorithm showed competitive search efficiency and superior scalability. We succeeded in designing large Silicon-Germanium (Si-Ge) alloy structures that Bayesian optimization could not deal with due to excessive computational cost. MDTS is available at https://github.com/tsudalab/MDTS.

Highlights

Complex materials design is a key topic in materials science and engineering
The difference between Bayesian optimization methods and traditional QSAR models is that the uncertainty of prediction is quantified as predictive variance: the candidates are scored by an acquisition function that takes into account both predicted merit and uncertainty
Bayesian optimization is very effective in finding optimal structures but has problems with scalability, as the acquisition function has to be applied to all candidates

Summary

Introduction

The design of a complex materials’ structure that meets certain criteria is often formulated as the problem of finding the optimal solution from a space of candidates [1,2]. To accelerate the materials design process, several experimental design algorithms have been used to find the optimal structure with as few experiments as possible (Figure 1). Bayesian optimization is very effective in finding optimal structures but has problems with scalability, as the acquisition function has to be applied to all candidates. Evolutionary algorithms such as genetic algorithms [10,11] are more scalable, but have many parameters, such as crossover and mutation rates, that must be tuned properly to obtain the best performance. In materials design, the amount of data available a priori is very limited, so tuning parameters using data may not be possible

Objectives

Methods

Conclusion