A chemically inspired convolutional neural network using electronic structure representation

• Published in: Journal of materials chemistry. A, Materials for energy and sustainability Materials for energy and sustainability, 2023-05, Vol.11 (19), p.1184-1194
• Main Authors: Mok, Dong Hyeon, Shin, Daeun, Na, Jonggeol, Back, Seoin
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Computational efficiency; Computer applications; Computing costs; Density of states; Electronic structure; Free energy; Heat of formation; Modules; Neural networks; Predictions; Representations; Screening; Stability analysis
• ISSN: 2050-7488; 2050-7496
• DOI: 10.1039/d3ta01767b

In recent years, the 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 the 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 the 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 an appropriate input for property prediction and the attention module is capable of properly featurizing DOS signals by capturing the contributions of each spin and orbital state. In addition, we statistically evaluated the stability screening performance of ESNet, measuring the 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. We developed Electronic Structure Network (ESNet) to predict formation energies using density of states extracted from initial structures. ESNet outperformed previously reported models that used other input features and architectures.