Density of States Prediction by a Metric-Optimized Equivariant Graph Neural Network

Density of states (DOS) is a crucial factor in elucidating the physical and functional properties of materials, playing a significant role in materials design. However, traditional density functional theory (DFT) calculations of DOSs are computationally expensive and inefficient, particularly when dealing with large structures and high-throughput electronic structure calculations. To address this challenge, we propose a metric-optimized equivariant graph neural network, called EqDOS, for DOS prediction, which can efficiently learn interatomic interactions from atomic position information. This capability allows the model to accurately predict DOS even with limited data, exhibiting excellent generalization capability. Additionally, by incorporating probability distribution metrics of high-order difference and Wasserstein distance for multiobjective learning, we effectively address the issues of oversmoothness and the sensitivity of sharp peaks in DOS curve prediction. These are especially important when handling materials with multiple compositions and structures in the periodic table. EqDOS enables the zero-shot prediction on the DOS profiles of ternary single-atom alloy surfaces with excellent performance outcomes, using the well-trained model based on the binary alloy surface data set. Furthermore, the predicted DOS by our model could serve as input for the downstream material performance prediction models. The results show that our model not only achieves higher accuracy in DOS curve fitting but also performs exceptionally well on the prediction of adsorption energy when combined with a downstream DOS Transformer for Adsorption (DOTA) model. This confirms that our model can accurately predict the DOS while preserving its physical significance. These findings not only validate the generalization and effectiveness of our model framework but also provide a new tool for the field of materials science to accelerate the discovery and optimization of new materials.