Speaker
Description
Predicting the optical properties of solids is of great interest for a wide variety of applications, such as solar cells, optical computing and sensors. However, calculating optical properties from first principles is computationally highly demanding. Recent advances have demonstrated that graph neural networks can accurately predict the optical spectra of a wide range of materials directly from their atomic configurations. Current models, however, are predominantly centered on predicting the frequency-dependent trace of the dielectric tensor or are limited by the amount of available training data. To address these limitations, we have calculated an extensive dataset of dielectric tensors for semiconductors and insulators, consisting of over 30,000 optical spectra, using automated density functional theory calculations. Using this dataset, we trained equivariant graph neural networks, achieving excellent prediction accuracy for the full dielectric tensor.