Machine Learning-Based Screening of Highly Stable and Active Ternary Pt Alloys for Oxygen Reduction Reaction
Joohwi Lee, Ryosuke Jinnouchi
Abstract
Alloying Pt with other chemical elements is a promising method for achieving higher catalytic activity for oxygen reduction reaction in proton-exchange membrane fuel cells. However, dissolution of secondary elements in acidic solutions is one of the major reasons for the poor durability of such alloy catalysts. Therefore, it is desirable to identify adequate compositions that can stabilize Pt alloys while maintaining high activity. First-principles calculations are a useful tool to search for an adequate alloy composition because it can predict the stability and catalytic activity based on kinetic models with reasonable accuracy; however, a high computational cost is unavoidable because an enormous number of atomic configurations need to be considered to compare the relative stabilities of the surface structures. In this study, we propose a rational and efficient screening strategy to find active and stable ternary Pt alloys from 4140 Pt15MmNn (m + n = 5) compositions with over two million surface structures. To realize efficient and accurate predictions of stability and activity, a new screening scheme combining crystal graph convolutional neural network-based machine learning (ML) method and the first-principles calculations is proposed. The ML model allows us to reduce candidate structures efficiently from two million to thousands. The first-principles screening of the suggested structures provides 29 ternary Pt alloys that have stable and active Pt-skin surfaces. The proposed method can be used to efficiently explore various catalytic materials that require millions of expensive first-principles calculations.