ИСТИНА

Stark broadening parameters play a crucial role in the interpretation of spectra from various plasma sources. Despite extensive experimental and theoretical efforts to retrieve Stark parameters and regular updates to existing databases, only a small fraction of known transitions are covered. Obtaining new values remains an important yet often difficult task. Experimental studies encounter limitations such as plasma source inhomogeneity, spectral interferences, insufficient resolution, and non-equilibrium particle interactions. Additionally, broadening of certain emission lines—such as those of multiply charged atomic ions — often cannot be observed in laboratory plasma. While quantum chemical calculations are resource-intensive, recent advances in machine learning methods have demonstrated high efficiency and accuracy across scientific domains, including spectroscopy. Given these considerations, we explore applicability of machine learning techniques for predicting Stark parameters and compare the obtained values with experimentally measured Stark parameters. We collected a dataset containing approximately 6500 values of experimentally measured Stark broadening and shift parameters. This dataset incorporates transitions of neutral atoms as well as of positive ions with charges up to 7 for 49 chemical elements. Inclusion of temperature as a feature to data representation allows us for prediction of temperature dependence of Stark broadening parameters. The optimized ensemble containing XGBoost, LightGBM and CatBoost models was trained on the dataset using data augmentation and standardization techniques adapted to specific features and imbalance of the data. It provides an average relative broadening prediction accuracy of ≈18% on an independent test set, that is comparable to the quality of the data used for training. Moreover, this ensemble demonstrates reasonable quality of prediction for chemical elements and ionization states that were excluded from training data. Since there is less data on Stark shift parameters than on broadening, we used the value of the broadening parameter as an additional feature to enhance predicting the shift parameters. We also compare the predicted values for some O II and N II lines with our experimental results, demonstrating benefits of having predicted Stark parameters for many transitions in fitting wide spectral ranges by a thermodynamic plasma model. Furthermore, weprovide access to the developed tool for predictions. The work was supported by the Fellowship from Non-commercial Foundation for the Advancement of Science and Education INTELLECT.