Physics guided deep learning generative models for crystal materials discovery

View on arXiv ← Back to list

Authors: Yong Zhao, Edirisuriya MD Siriwardane, Jianjun Hu

Published: 2021-12-07 06:54:48+00:00

AI Summary

This paper presents a physics-guided deep generative model for creating crystal structures. The model incorporates physically oriented data augmentation, loss function terms, and post-processing to generate crystal structures with higher physical feasibility, expanding beyond previously limited cubic structures.

Abstract

Deep learning based generative models such as deepfake have been able to generate amazing images and videos. However, these models may need significant transformation when applied to generate crystal materials structures in which the building blocks, the physical atoms are very different from the pixels. Naively transferred generative models tend to generate a large portion of physically infeasible crystal structures that are not stable or synthesizable. Herein we show that by exploiting and adding physically oriented data augmentation, loss function terms, and post processing, our deep adversarial network (GAN) based generative models can now generate crystal structures with higher physical feasibility and expand our previous models which can only create cubic structures.

Key findings

The physics-guided enhancements significantly improved the model's ability to generate physically feasible crystal structures. The model successfully generated structures for 12 different space groups, including non-cubic structures, expanding upon previous work limited to cubic structures. DFT calculations validated the dynamic stability of some generated structures.

Approach

The authors use a Wasserstein Generative Adversarial Network (GAN) enhanced with physics-guided components. These include self-augmentation of atom sites using symmetry operations, physics-guided loss functions to prevent atom collisions, and post-processing to cluster and merge overlapping atoms.

Datasets

Training crystal structures were collected from the Materials Project, focusing on ternary materials with three base atom sites and twelve space groups.

Model(s)

Wasserstein Generative Adversarial Network (GAN)

Author countries

USA

← Previous