Comparison of dimensionality reduction algorithms for solving the inverse problem of fluorescence spectroscopy using machine learning methods
Abramov I. S.1, Isaev I. V.2, Dolenko S. A.2; 1 Faculty of Physics, M. V. Lomonosov Moscow State University; 2 D. V. Skobeltsyn Institute of Nuclear Physics, M. V. Lomonosov Moscow State University
Abstract
One of the major ecological challenges is the release of toxic substances, particularly heavy metal ions, into the environment. Their concentration increases due to both natural processes and anthropogenic activities, making the quantification of these ions in water and soil a critical task for environmental monitoring. This work addresses the inverse problem of fluorescence spectroscopy for determining ion concentrations in solutions using their 2D excitation-emission matrix (EEM) spectra. Currently, there are no universal models capable of accurately correlating spectral features with element concentrations in multicomponent solutions. Machine learning methods trained on experimental data offer an effective solution to this problem.
However, 2D EEM spectra are high-dimensional and often contain redundant or uninformative data. For instance, adjacent spectral channels may exhibit strong correlations. Direct use of all spectral channels not only increases computational costs but may also lead to model overfitting.
We compare the performance of various dimensionality reduction algorithms applied to 2D fluorescence EEM spectra and evaluate their effectiveness in solving the inverse problem of determination of ion concentrations.
Speaker
Abramov Ivan
Faculty of Physics, M. V. Lomonosov Moscow State University
Russia
Discussion
Ask question
