![]() |
ИСТИНА |
Войти в систему Регистрация |
ФНКЦ РР |
||
Controlling the ionic composition of process waters requires remote and express methods. Optical spectroscopy is the most promising technique in this regard. Carbon dots (CD) have broad prospects of application as optical fluorescent nanosensors in the tasks of media diagnostics. However, the lack of analytical models that adequately describe the fluorescent response of CD to environmental changes hamper the analysis of CD fluorescence spectra in media in the presence of several ions. Therefore, to create a CD-based nanosensor of ions it is necessary to use machine learning (ML). Several challenges must be faced to obtain representative sets of spectroscopic data required to train ML models. Unfortunately, many strategies developed for increasing the representativity of datasets are unsuitable for specific spectroscopic data. In this study, a comparative analysis of two approaches to solving the inverse problem of fluorescence spectroscopy for simultaneous determination of the concentration of heavy metal ions (Cu2+, Ni2+, Cr3+) and nitrate ions in water by CD FL spectra is carried out. These approaches included: 1) training of convolutional neural networks (CNN) on experimentally obtained fluorescence excitation-emission matrices (EEM); 2) using variational autoencoders (VAE) to increase the representativity of the training dataset with subsequent training of CNN. The use of the second approach enabled reduction of mean absolute error in determining the ions concentration by 60% for Cr3+, by 41% for Ni2+, by 62% for Cu2+, by 48% for NO3-. This study has been conducted at the expense of the grant of the Russian Science Foundation № 22-12-00138, https://rscf.ru/en/project/22-12-00138/.