Comparing approaches to improving representativity of spectroscopic data using variational autoencoders

Solving inverse problems in optical spectroscopy of multicomponent solutions to determine component concentrations is a complex task. One effective approach to solving this problem is the use of artificial neural networks. However, one of the major challenges in this approach is the limited representativity of experimental data, due to the complexity and high cost of large-scale physical experiments. In this paper, algorithms for generating additional model data using variational autoencoders are explored and compared to enhance the representativity of the training dataset. The results show that the most promising approach is the use of a standard (unconditioned) variational autoencoder, generating patterns from the uniform distribution in the latent space. Further research should focus on identifying the optimal distribution in the latent space for generating patterns.