CrysGraphFormer: an equivariant graph transformer for prediction of lattice thermal conductivity with interpretability

Abstract

To address the challenges of high error rates and poor generalization in current deep learning models for predicting lattice thermal conductivity (LTC), we introduce CrysGraphFormer, an innovative equivariant crystal graph transformer model tailored for this task. The model incorporates an improved multi-head self-attention mechanism and human-designed feature descriptors. By utilizing a message-passing mechanism to update node information, it introduces relative coordinate differences to represent crystal symmetry, avoiding the traditionally used complex and computationally expensive higher-order representations. We constructed a comprehensive dataset containing 5729 LTC data points (300 K), including 5477 materials from AFLOW, 112 MAX and MAB phase materials calculated using VASP, and 140 for half-Heusler alloys. Experimental results demonstrate that the CrysGraphFormer model achieves state-of-the-art performance in LTC prediction tasks and excels in predicting fundamental properties. The model offers good interpretability, providing insights from chemical and materials science perspectives. Furthermore, we validated the model's application potential in the field of thermoelectric materials by predicting the LTC of 59 thermoelectric materials and 55 ternary semiconductor materials, with results consistent with DFT calculations. Finally, the uncertainty of CrysGraphFormer was assessed using the Monte Carlo Dropout method.