ИСТИНА

Stark broadening parameters are crucial for analysis and modeling spectra from various emission sources. For instance, the lack of known Stark parameters for atomic lines in white dwarf spectra hinders accurate analysis and interpretation. Despite decades of experimental and theoretical studies to retrieve Stark parameters and regular additions to existing databases, only a small percentage of known transitions are covered. Obtaining new values remains an important and often challenging task. Experimental studies face limitations such as plasma source inhomogeneity, spectral interferences, insufficient spectral resolution, and non-equilibrium particle interactions. Broadening of some emission lines like lines of doubly charged atomic ions cannot be observed in laboratory plasma due to low Stark broadening values. Quantum chemical calculations are limited by high computational costs and strong electronic state interactions or relativistic effects for heavy atoms. However, rapid advancements in machine learning (ML) methods have resulted in their high efficiency and accuracy in many scientific areas including spectroscopy. Artificial neural networks (ANNs) have been successfully implemented for tasks ranging from hyperspectral image analysis to predicting reaction rates or substance properties based on its structure. These considerations led us to the application of ML for predicting Stark parameters. We created a table format to represent electronic level configurations, terms and energies for each atomic or ionic transition. Our database contains approximately 1000 atomic and ionic transitions with experimental Stark parameter values. The availability of parameters measured at different temperatures allows us to add a temperature parameter to the database and predict temperature-dependent Stark broadening parameters using ML methods. We trained and tested several classical machine learning models (KNN, Random Forest, Gradient Boosting) and ANNs (TabNet and custom MLP) for predicting Stark broadening and shift parameters and tested the results on independent subsets. Boosting algorithms demonstrated the highest accuracy after implementation of data scaling and augmentation procedures. ANNs were not yet competitive with classical ML models but their accuracy strongly depends on the size of the training dataset. Adding new experimental values to the database in future could improve their accuracy. We also tested model generalization by evaluating them on an independent subset of transitions for three chemical elements not seen in the training set. We added predicted Stark parameters for a large number of transitions to a thermodynamic model simulating low-temperature plasma emission [1]. Comparing experimental and simulated spectra with predicted Stark parameters shows that having parameters for many transitions benefits spectra interpretation and plasma diagnostics. ACKNOWLEDGEMENTS The work was supported by the Fellowship from Non-commercial Foundation for the Advancement of Science and Education INTELLECT.