Develops ESNet, an Artificial Intelligence Model for Predicting Steady-state Energy

▲ (from left) Professor Back Seo-in and graduate student Mok Dong-hyeon of the Department of Chemical and Biomolecular Engineering

Professor Back Seo-in’s research team in the Department of Chemical and Biomolecular Engineering developed ESNet (Electronic Structure Network), an artificial model that predicts the energy of a stable ground state from an initial crystal structure, collaborating with Professor Na Jong-geol’s research team from Ewha Womans University.

The research results were published in the Journal of Materials Chemistry A (citation index 14.511, JCR top 7.14%), an international journal in the field of materials design, under the title 'A Chemically Inspired Convolutional Neural Network Using Electronic Structure Representation.'

High-throughput virtual screening is a method for discovering new materials by randomly generating a large number of crystal structures and exploring them to quickly find those meeting a set of desired properties. In order to perform efficient screening, it is critical to achieve quick and accurate evaluations of the properties of crystal structures generated in the process.

AI models can predict the properties of crystal structures in less than a second, and there have been recent attempts to reduce the duration required for searches by replacing the time-consuming electronic density functional theory calculations with AI models to determine the properties of the crystal structures.

However, existing AI models are either too sensitive in their crystal structure representation, only predicting the properties of a given crystal structure as it is, or too insensitive to distinguish structural differences between polymorphs. As such, the AI models show very poor performance in predicting the properties of a crystal structure in its ground state based only on a randomly generated crystal structure (hereinafter referred to as the initial structure). As a result, the models find it hard to perform high-speed discovery for material development, given that additional structure optimization is required to obtain the ground-state crystal structure.

The research team proposed the density of states as a crystal structure representation method that can overcome these limitations. The density of states has different values for different crystal structures, and it shows appropriate structure sensitivity, as the structural information does not change significantly before or after structure optimization. Based on this fact, the team developed ESNet as an artificial intelligence model that can predict the energy of the ground state from the density of states of the initial structure.

ESNet is based on a convolutional neural network applied with a self-attention module to capture important information for intensive learning, and it learns the density of states after dividing them into orbital and spin states, using a structure based on domain knowledge that integrates the density of states. Therefore, information about the stable crystal structure can be efficiently extracted from the density of states of the initial structure without the structure-optimization process.

ESNet has outperformed other artificial intelligence models developed for the same purpose in terms of prediction accuracy and was found to be able to identify thermodynamically stable materials with only 18% of the computational effort compared to existing models. This study is meaningful in that it has developed a source technology that combines artificial intelligence and chemical simulation to create optimal materials without physical experiments.

The research was performed as a collaborative effort between Sogang’s Professor Back Seo-in and graduate student Mok Dong-hyeon (co-first author) from the Department of Chemical and Biomolecular Engineering and Ewha Womans University’s Professor Na Jong-geol (corresponding author) and graduate student Shin Da-eun (co-first author), receiving support from the 'Young Researcher Program,' 'C1 Gas Refinery R&D Center,' 'Priority Research Institute Program' of the National Research Foundation of Korea, and KISTI National Supercomputing Center.

▶ Read the article: 

https://pubs.rsc.org/en/content/articlelanding/2023/TA/D3TA01767B

▲ (Clockwise from top-right) Professor Back Seo-in and graduate student Mok Dong-hyeon of the Department of Chemical and Biomolecular Engineering at Sogang University, and graduate student Shin Da-eun and Professor Na Jong-geol of the Department of Chemical Engineering and Materials Science at Ewha Womans University