Star-gen: An HPC-AI Ensemble Framework for Boosting the whole process of Materials Design

Artificial intelligence (AI) has emerged as a promisingly new paradigm for scientific research. Most of the emerging applications are driven by HPC-AI ensembles and trained on massive amount of data generated by experiments or simulations. For the purpose of promoting the effectiveness and efficiency, it is vital to develop an orchestration framework to couple data processing, simulation and training AI models. In this presentation, we are proposing Star-gen, an HPC-AI framework that tightly couples essential components with enhanced task-level interactive mechanisms. To improve the quality of simulated scientific data, Star-gen constructs a large-scale density functional theory (DFT) computational database for 76,000 materials using up to 48,000 physical cores on Tianhe-2 supercomputer, consuming a total of about 100 million CPU hours. Based on the data, five representative graph neural network (GNN) models achieved significant improvement in accuracy when predicting material characteristics compared to previous studies. By adopting the framework, domain scientists can design a novel bimetallic nanoparticle and discover two potential drugs for ameliorating metabolic complications induced by COVID-19. The data repositories, AI models, and data processing tools in Star-gen have been released and made accessible with support for thousands of registered users by now.