Prediction of Stark Broadening of Atomic Spectral Lines Using Machine Learning

Abstract

Stark broadening and shift of atomic spectral lines are fundamental parameters for the diagnosis of astrophysical and laboratory plasmas. Traditional theoretical calculation methods, such as Semi-Classical Perturbation (SCP) or Modified Semi-Empirical (MSE) approaches, are computationally expensive and often limited in coverage. This paper presents a data-driven approach based on machine learning to achieve rapid and high-precision prediction of Stark parameters using large-scale data from the STARK-B database. The core innovation of this study lies in the development of a deep feature engineering pipeline integrated with physical prior knowledge. Addressing the complex textual representations in raw spectroscopic data, I developed specialized parsing algorithms to extract physically meaningful numerical features—such as orbital occupancy counts, multiplicity, L-quantum numbers, and parity—from electron configurations, spectral terms, and total angular momentum (J). Over 130,000 processed data points were used to evaluate the performance of various models, including Random Forest, Neural Network, and XGBoost. The final model demonstrated exceptional accuracy for Stark widths, achieving a Coefficient of Determination (R^2) of 0.952 and a Median Relative Error (MdRE) of just 5.71% on an independent test set. While Stark shift predictions proved more challenging due to near-zero values, the model successfully captured macroscopic trends (R^2\approx0.92). This research confirms that machine learning, when combined with deep physical feature extraction, offers a scalable and reliable alternative for populating large-scale atomic spectral databases.