Prediction of electron-solid interaction parameters using machine learning

Area of Science:

• Materials Science
• Physics
• Computational Science

Background:

• Electron backscattering coefficient and electron stopping power are crucial for radiation, materials science, semiconductor manufacturing, and space exploration.
• Accurate data is vital for calculations, simulations, and advancing scientific understanding and safety.
• Machine learning (ML) offers a promising approach to improve data quality and completeness.

Purpose of the Study:

• To develop a stacking ensemble machine learning (EML) technique for generating electron-solid interaction parameters.
• To predict electron backscattering coefficient and electron stopping power for any material across a wide energy range.
• To utilize fundamental material properties as input for the EML model.

Main Methods:

• A stacking ensemble ML model was constructed using base learners (Bagging Regressor, k-NN, Random Forest, SVR, XGBoost) and a meta-learner.
• Two public databases with 4030 data points were used for training and testing.
• Model performance was evaluated using R-squared, MAE, RMSE, and MAPE metrics.

Main Results:

• The ensemble model combining Random Forest and XGBoost with a k-NN meta-learner demonstrated superior performance.
• Error metrics indicated a close fit to training data and accurate predictions on unseen test data.
• The model successfully estimated new backscattering and stopping power data.

Conclusions:

• The developed ML model achieves high prediction accuracy for electron interaction parameters across various materials and energies.
• The study highlights the efficacy of ML in addressing complex physics challenges.
• The findings facilitate data discovery and support advancements in related scientific and technological fields.