근사적 등변 프레임 평균화를 통한 결정질 재료용 CGCNN 향상

초록

Deep learning-based computational modeling of materials data has gained attention for offering significantly higher computational efficiency and competitive accuracy compared to ab initio methods. CGCNN, a graph neural network model with a simple and extensible architecture, has shown promising performance; however, its reliance solely on interatomic distances limits its ability to capture full three-Dimensional (3D) geometric structures, resulting in constraints on prediction accuracy. In this study, we propose Frame Averaging-based CGCNN (FACGCNN), an extension of CGCNN that incorporates atomic coordinates and relative positions into node and edge embeddings, respectively. To address the issue of sensitivity to coordinate system transformations we apply an approximate equivariant frame averaging technique. Experiments conducted on the Carolina dataset, which consists of inorganic crystal structures, demonstrate that the proposed method outperforms the original CGCNN in terms of predictive accuracy.

키워드