POAT: Material property prediction from limited data via Pointwise Distance Distribution and Offset Attention

The properties of the material determine the various applications of the material. In the last decades, material properties have often been determined through trial-and-error experiments, which are slow and costly, or Density Functional Theory (DFT) calculations, which are computationally intensive and limited in terms of material structure. Recently, an encoding method called Pointwise distance distribution (PDD) has achieved impressive results in representing crystal structures. However, this method is unable to effectively deal with the periodicity and denseness problems of crystals with complex structures, which is not conducive to predicting material properties. In this paper, we propose a transformer model based on Pointwise distance distribution encoding and Offset ATtention mechanism (POAT). Our network can represent crystal structures in a flexible manner and effectively handle crystal periodicity and denseness problems. Numerical experiments on the JARVIS-DFT and MatBench structure datasets show that the proposed model achieves the state-of-the-art performance in most of the property prediction tasks, particularly demonstrating superior robustness when the training data is limited. The POAT model also shows significant efficiency advantages in training and prediction time compared to graph network models. An ablation study further investigates the importance of the offset attention mechanism in the POAT model. Furthermore, the developed model is applied to predict the heat capacity, further illustrating the versatility of the model.