Chemically Inspired Convolutional Neural Network using Electronic Structure Representation

In recent years, a development of appropriate crystal representations for accurate prediction of inorganic crystal properties has been considered as one of the essential tasks to accelerate materials discovery through high-throughput virtual screening (HTVS). However, many of them were developed aiming to predict properties of the given structures, although property predictions of ground state structures using unrelaxed structures as inputs are much more important in practical HTVS. To tackle this challenge, we develop a chemically inspired convolutional neural network based on convolution block attention modules using density of states of unrelaxed initial structures (IS-DOS) as inputs. Our model, Electronic Structure Network (ESNet), achieved the highest accuracy for predicting formation energy, proving that IS-DOS is appropriate input for the property prediction and the attention module is capable of properly featurizing DOS signals by capturing contributions of each spin and orbital state. In addition, we statistically evaluated a stability screening performance of ESNet, measuring computational cost and capability of materials discovery simultaneously. We found that ESNet outperformed previously reported models and various models with different types of input features and architectures. Indeed, ESNet successfully discovered 926 stable materials from 15,318 unrelaxed structures with 82 % reduced computational cost compared to the complete DFT validation.