Screening for wide-gap, high-refractive-index materials via frequency-dependent dielectric function prediction aided by convolutional autoencoders

In this work, we construct a database containing approximately 30 000 entries for the frequency-dependent dielectric function within the independent-particle approximation, covering a wide range of semiconductors. Calculations are performed using the PBE and TASK approximations to the exchange-correlation potential. Building on these results, we evaluate various approaches to dimensionality reduction of the spectra, including convolutional autoencoders, which achieve effective compression to a latent space around 50 times smaller. These models are comparatively inexpensive to train and help reduce the computational cost associated with training large, off-the-shelf neural network models. The resulting models are subsequently employed to screen for wide-bandgap, high-refractive-index materials.