Fig. 1. Schematic visualization of the 3 types of learning models for optical properties of materials. The first algorithm (top) illustrates the variational autoencoder (VAE) that autoencodes images Ĩi from Iivia a latent space representation Z[combining tilde]i. The encoder performing the mapping Ii to Z[combining tilde]i is called EVAE, the decoder performing the mapping from Z[combining tilde]i to Ĩi is called DVAE. The second model employs the latent space of the VAE but decodes Z[combining tilde]i into an absorption spectrum S[combining tilde]i (instead of an image) using a deep neural net (DNN), producing an image to spectrum prediction model. The cVAE reconstructs images Ĩi from images Ii and spectra Si such that the latent space vector Z[combining tilde]i encodes image and spectral information, which is decoded in conjunction with a specific absorption spectrum Si to yield an image that is predicted to exhibit the specified absorption properties.
