CC BY
0Determination of heavy metal ions in river water by spectroscopy and machine learning: use of transfer learning approach
A. Guskov12, I. Isaev13, S. Burikov12, T. Dolenko12, K. Laptinskiy1, S. Dolenko1*
1-D. V. Skobeltsyn Institute of Nuclear Physics, M.V. Lomonosov Moscow State University, 1/2 Leninskie Gory,
Moscow, 119991 Russian Federation
2- Faculty ofPhysics, M.V. Lomonosov Moscow State University, 1/2 Leninskie Gory,
Moscow, 119991 Russian Federation
3- V.A. Kotelnikov Institute of Radio Engineering and Electronics, 11/7 Mokhovaya st.,
Moscow, 125009 Russian Federation
* dolenko@srd.sinp.msu.ru
Heavy metals ions are a class of substances that are often present not only in wastewater, but also in nature waters, and that are harmful for public health. That is why detection of such impurities and determination of their concentrations is a very topical task. Use of methods of chemical analysis provides high precision of measurement, but it requires special sample preparation, expensive reagents and high qualification of personnel. At the same time, spectroscopic methods are express and they can be used in remote mode.
However, to provide sufficient precision of spectroscopy-based measurements, one should use special data processing methods. In this study we use laser Raman spectroscopy, IR and optical absorption spectroscopy. While the shape of water spectra of all the three types is highly sensitive to presence of ions, simultaneous presence of several types of ions causes specific non-linear dependencies in spectra channel intensities on ion concentrations. Moreover, adequate modeling of spectra of such solutions is yet far beyond reasonable computational capabilities. By all these reasons, machine learning (ML) methods turn out to be a class of data processing methods that are very effective in solving such inverse problems.
At the same time, use of ML methods encounters its own specific difficulties. To provide an acceptable precision of concentration measurements, one needs a representative set of data to train the ML methods. As specified above, model spectra are unavailable, and obtaining a large enough database of experimental spectra is difficult and may be quite laborious and expensive.
If we speak of nature waters, there is one more problem complicating the spectra analysis. Nature waters always contain dissolved organic matter (DOM), which has intense and highly variable fluorescence partly overlapping the most interesting spectral regions. As DOM fluorescence depends on the origin of water, there is a need to introduce a special procedure that would make ML models resilient to variations in nature water spectra.
In this study, we consider two approaches that turn out to be efficient in this situation. One is integration of several spectroscopic methods - joint use of several types of spectra of the same sample. The other one is use of the so called transfer learning approach. Within this approach, one first trains an ML model on a large enough and representative dataset obtained in simple conditions (here we use spectra of about 3700 model solutions prepared in laboratory conditions in distilled water). After that, a much smaller amount of patterns (spectra) obtained for a specific situation (in this case, 200-400 samples prepared in river water) is used for fine tuning or the ML model, providing the sufficiently low error of determination of ion concentrations. We demonstrate [1] that both approaches are effective, and in some situations the most effective way is to use both method integration and transfer learning at once.
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/.
[1] A.A. Guskov, I.V. Isaev, S.A. Burikov, T.A. Dolenko, K.A. Laptinskiy, S.A. Dolenko. Transfer Learning for Neural Network Solution of an Inverse Problem in Optical Spectroscopy. Moscow University Physics Bulletin, V. 78, Suppl. 1, pp. S115-S121 (2023).
