ИСТИНА

This study constitutes an investigation into the potential improvement of neural network methods through the augmentation of the training dataset using a variational autoencoder (VAE). We consider the inverse problem of spectroscopy of multi-component water solutions, aimed at determining the concentrations of various ions in the solutions based on their spectral data (Raman, IR or optical absorption spectroscopy). While the shape of the spectra is sensitive to the concentrations of ions, the dependence of spectral intensities on ion concentrations in multi-component solutions is complex and non-linear, thus requiring analysis of many spectral channels at once. Such analysis may be performed using machine learning methods, e.g. neural networks. However, to train a neural network, a large dataset is required. Adequate modeling of spectra of multi-component solutions is yet far beyond reasonable computational capabilities. The required dataset may be obtained through laboratory measurements, but such experiment is very laborious and expensive. Furthermore, the experimental data may also exhibit a significant imbalance in the space of target values that can interfere with the training of a regression neural network model. Thus, the issue of adequately expanding the training dataset is highly pertinent. This study examines the possibility of augmenting the training dataset by generating extra spectra using a VAE. The aim is to provide reduction of the error of solving the inverse problem. There are several possible approaches to such generation. It seems that the simplest way is to use a conditioned VAE (cVAE) trained on experimental data. However, we demonstrate that the quality of such generation is insufficient – generated spectra differ too much from experimental spectra with the same ion concentrations. Expanding the experimental dataset with such generated spectra even increases the error of determination of ion concentrations. Two other approaches use VAE. In this case, we require a way to determine the target ion concentrations for each of the generated spectra. This may be done by some ML regression model trained on experimental data – either in the feature space of the spectra or in the latent space of VAE. Subsequently generated spectra can be used in various ways along with experimental ones during the training of regression neural networks solving the inverse problem. In this study, we compare these two approaches and discuss their merits and shortcomings. The impact of the level of noise in the experimental data, and of the level of noise added to experimental spectra during training of the VAE, on the generated spectra is also explored. Possible ways of improving data distribution during spectra generation are also discussed. The study was carried out at the expense of the grant No. 24-11-00266 from the Russian Science Foundation, https://rscf.ru/en/project/24-11-00266/.