Fully automated high-throughput computer-based catalytic material screening framework and its application on the new-generation Tianhe supercomputer

Abstract

The integration of high-performance computing with machine learning (ML) has established a transformative scientific paradigm that significantly enhances the efficiency of material discovery, particularly in the search for catalysts in alternative energy research. However, significant challenges remain in the utilization of available computational resources to accelerate the screening of catalyst materials. In this study, we implement a high-throughput framework on the new-generation Tianhe supercomputer, featuring the development of a Ping-Fault Recovery algorithm, single-task optimization for Density Functional Theory (DFT) to maximize efficiency, and enhanced task scheduling using a two-level scheduling strategy to ensure efficient utilization of the abundant computational resources of the supercomputer. This framework facilitates the identification of 2,028 candidate surfaces across 868 intermetallics from 2,713,897 unique adsorption sites, achieving a screening speed 193 times faster than traditional methods. Alloys composed of Mo, Nb, and V are used as case studies to provide a detailed elucidation of the process of identifying the most effective catalytic surfaces. The framework achieved the best single-day candidate hit performance on 18,106 nodes, completing in one day what previously took a year. This supercomputer-based framework optimizes the use of computational resources, driving innovation in catalyst material discovery.