High-accuracy prediction of Kβ/Kα intensity ratios via explainable stacked-ensemble learning: A web-based...

Area of Science:

• Atomic Physics and Spectroscopy
• Computational Physics
• Data Science in Science

Background:

• Kβ/Kα intensity ratios are crucial for understanding atomic shell transitions and radiative probabilities.
• Traditional theoretical models have limitations in accurately predicting these ratios.
• A comprehensive dataset of experimental measurements is essential for developing predictive models.

Purpose of the Study:

• To systematically evaluate machine learning (ML) algorithms for predicting Kβ/Kα intensity ratios.
• To compare the performance of ML models against traditional theoretical approaches like Scofield calculations.
• To develop a practical tool for predicting Kβ/Kα ratios based on experimental parameters.

Main Methods:

• Compiled a dataset of 2124 experimental Kβ/Kα ratios (Z=11-96).
• Utilized multivariate imputation by chained equations (MICE) for missing data and one-hot encoding for categorical variables.
• Trained and evaluated seven individual ML models and seven stacking ensemble models using 10x10-fold cross-validation.

Main Results:

• A stacked ensemble model with a BRNN meta-learner achieved the highest accuracy (RMSE=0.009, R²=0.973).
• The best ML model reduced test error by ~48% compared to Scofield theory, with significant Diebold-Mariano test results (p<0.001).
• SHAP analysis identified atomic number as the primary predictor, with sample purity and excitation source as secondary factors.

Conclusions:

• Data-driven ML approaches offer superior accuracy and interpretability for predicting Kβ/Kα ratios compared to classical theories.
• An online R/Shiny calculator was developed for practical, real-time Kβ/Kα predictions.
• This work advances atomic databases, detector calibration, and understanding of high-Z elements.