Equivariant graph neural networks outperform existing models in predicting optical spectra for materials screening. The adapted GotenNet achieves superior performance, especially in the 0-8 eV range and for static real permittivity prediction, critical for thin-film optics.