SFTGNN: An Efficient Spherical Fourier Transform-Enhanced Graph Neural Network for High-Accuracy Crystal Property Prediction

Crystalline materials, characterized by periodic atomic arrangements, play a fundamental role in materials science due to the strong correlation between their structure and physical properties. While graph neural networks (GNNs) have shown promise in predicting crystal properties, many existing models either neglect crucial angular features such as bond angles or suffer from high computational costs, as seen in methods like ALIGNN with <i>O</i>(<i>k</i> ²) complexity for local angle calculations. In this work, we introduce SFTGNN, a novel architecture that integrates spherical Fourier transforms (SFT) with GNNs to efficiently incorporate three-dimensional geometric information into crystal property prediction. By projecting atomic local environments into the spherical harmonic domain, SFTGNN captures angular dependencies without the need to explicitly enumerate all bond angles, while reducing the computational complexity of bond angle encoding from <i>O</i>(<i>k</i> ²) to <i>O</i>(<i>k</i>) and obviating the requirement for additional graph construction. We have validated the effectiveness of this method and evaluated the predictive performance of the model on the Materials Project and Jarvis data sets. SFTGNN achieves state-of-the-art results in multiple crystal property prediction tasks, while training 5.3 times faster than ALIGNN. These results highlight the potential of combining spatial geometric representations with graph-based learning for efficient and accurate modeling of crystal structures.