PSCG-Net: A Multiscale Crystal Graph Neural Network for Accelerated Materials Discovery.

Abstract

The discovery of new materials is crucial for progress in energy, electronics, and sustainable technology. Traditional machine learning approaches, including graph neural networks (GNNs), often fall short because they cannot capture long-range interactions in crystalline materials due to fixed cutoff radii. To overcome this limitation, a pair-scalable crystal graph neural network (PSCG-Net) is proposed. This framework incorporates multiscale structural representations inspired by the pair distribution function and uses graphs with various cutoff distances to account for both short-range and long-range atomic interactions. Tested on over 150,000 crystal structures, PSCG-Net outperforms the baseline Crystal Graph Convolutional Neural Network model by achieving a mean absolute error of 0.065 eV in formation energy prediction. The model's effectiveness is further supported by consistent results across six diverse data sets and confirmed by first-principles calculations using hybrid functionals in band gap-type predictions. Additionally, PSCG-Net is demonstrated to be practical for screening high-performance materials in photovoltaics, dielectrics, and superconductors. By accurately capturing hierarchical atomic interactions, this approach accelerates the design and discovery of materials and offers a versatile framework applicable to multiscale challenges in various scientific disciplines. This framework not only enhances predictive accuracy but also paves the way for breakthroughs in materials science research and technological innovation.