Accurate prediction of tensorial spectra using equivariant graph neural network

Optical spectroscopies provide a powerful means to probe light-matter interactions and complex electronic features that are crucial for the development and optimization of optoelectronic devices, where performance is closely tied to the underlying electronic spectrum. However, realistic modeling of tensor optical responses in materials remains computationally demanding and challenging. Here we introduce the Tensorial Spectra Equivariant Neural Network (TSENN), an equivariant graph neural network architecture that maps crystal structures directly to their full photon-frequency-dependent optical tensors. By encoding isotropic sequential scalar components and anisotropic sequential tensor components into spherical tensor representations, TSENN ensures symmetry-aware predictions consistent with crystalline symmetry constraints. Trained on frequency-dependent dielectric tensors of 1,432 bulk semiconductors, the model achieves a mean absolute error of 0.127, demonstrating its potential for efficient and general modeling of optical properties. Our framework opens new avenues for data-driven design of anisotropic optical responses to accelerate materials discovery for optoelectronic applications. Modeling full tensorial optical responses of crystalline materials is computationally demanding. Here, the authors introduce TSENN, an equivariant graph neural network that predicts symmetry-consistent, frequency-dependent optical tensors directly from crystal structures.