Equivariant Graph Neural Networks Improve Optical Spectra Prediction for Materials Screening

2026-06-17

AI Summary

This work addresses the limitations of existing optical spectrum prediction methods, which suffer from low-accuracy theoretical calculations or rotationally invariant scalar descriptors that inadequately capture material geometric structures, thereby hindering efficient optoelectronic material screening. For the first time, we apply an equivariant graph neural network based on the GOTENN architecture to high-accuracy optical spectrum prediction. Trained on a large-scale dataset comprising 10,533 material structures with spectra generated via the random phase approximation (RPA), the model significantly outperforms state-of-the-art approaches in predicting absorption spectra across the 0–8 eV energy range and the static real dielectric constant. By overcoming the bottlenecks of conventional feature representations, this method provides an effective tool for high-throughput screening of thin-film optical materials.

Abstract

Scalable prediction of optical spectra is a critical component of high-throughput materials screening for optoelectronic applications such as solar cells. Existing surrogate models are trained on spectra computed from lower levels of theory or rely on rotation-invariant scalar features, limiting their geometric expressiveness. We explore the use of equivariant graph neural networks for optical spectra prediction, adapting GotenNet to this task and evaluating it on multiple datasets including a recently published collection of 10,533 structures with spectra computed at the level of the random phase approximation (RPA). The proposed model outperforms the current state of the art, with the largest gains in the 0-8 eV range and on predicting the static real permittivity, both of particular relevance for thin-film optics.

Problem

• optical spectra prediction
• materials screening
• geometric expressiveness
• equivariant graph neural networks
• random phase approximation

Innovation

• equivariant graph neural networks
• optical spectra prediction
• materials screening
• GotenNet
• geometric expressiveness