Advancing Novel Material Design by “Supercomputer + AI”

Pin Chen1, Hui Yan, Qing Mo, Zexin Xu

National Supercomputer Center in Guangzhou, School of Data and Computer Science, Sun Yat-sen University, Guangzhou, 510006, China

EXTENDED ABSTRACT: Recently, due to the rapid development of supercomputers and software parallel technology, it has become possible to quickly generate large-scale computational material data through high-throughput computations. And allowing scientists to screen materials based on specified properties, or use artificial intelligence (AI) technology to predict material properties and design new materials. In this report, we have developed a data-driven material design method by combining supercomputer and AI technology. We applied for more than 48,000 CPU cores on the "Tianhe-2" supercomputer to perform high-throughput computations by PBE method, and constructed a large-scale computational database. The computational material structures are all collected from experimentally-observation database. Based on the computational data, we built a graph convolutional neural (GCN) network model that accurately predicted the material properties, and solved problem of poor efficiency in obtaining material properties. We further employed a small amount of experimental data points to fine tune the model through the transfer learning method. Our new model achieved a higher accuracy, and alleviated accuracy problem caused by PBE computational method. We applied the model to predict physical properties of the hypothetical materials, and obtained more accurate results than the PBE calculation method, which has the potential to design new materials. In conclusion, we presented a data-driven model by leveraging large-scale computation data, AI and small-scale experimental data to predict material properties. Our method is also applicable to other materials. The relevant data, AI models and software codes in this report are all available at Matgen platform (https://matgen.nscc-gz.cn).