A Universal Machine Learning Framework for Electrocatalyst Innovation: A Case Study of Discovering Alloys for Hydrogen Evolution Reaction

Massive efforts have been made to develop efficient electrocatalysts for green hydrogen production. The introduction of machine learning (ML) has brought new opportunities to the design of electrocatalysts. However, current ML studies have shown that the efficiency and accuracy of this method in electrocatalyst development are severely hindered by two major problems, high computational cost paid for electronic or geometrical structures with high accuracy, and large errors resulted from those easily accessible and relatively simple physical and chemical properties with lower level of accuracy. Here, a universal ML framework is proposed that achieves local structure optimization by using local machine learning potential (MLP) to efficiently obtain accurate structure descriptors, and by combining simple physical properties with graph convolutional neural networks, 43 high-performance alloys are successfully screened as potential hydrogen evolution reaction electrocatalysts from 2973 candidates. More importantly, part of the best candidates identified from this framework have been verified in experiments, and one of them (AgPd) is systematically investigated by ab initio calculations under realistic electrocatalytic environments to further demonstrate the accuracy. More significantly, the computational efficiency and accuracy can be compromised with this local MLP optimized structural descriptor as the input, and a new paradigm could be established in designing high-performance electrocatalysts.

Automated summary

Efforts have been made to develop efficient electrocatalysts for green hydrogen production, and the introduction of machine learning has brought new opportunities. However, current studies show that the efficiency and accuracy of machine learning in electrocatalyst development are hindered by computational cost and errors from simple physical and chemical properties. This study proposes a universal machine learning framework that combines local machine learning potential (MLP) and graph convolutional neural networks to screen and optimize high-performance electrocatalysts.