Ensemble machine learning methods: predicting electron stopping powers from a small experimental database.

Area of Science:

• Materials Science
• Computational Physics
• Nuclear Science

Background:

• Electron stopping power (SP) is crucial for microanalysis, dosimetry, and detector design.
• Experimental SP data is limited to a few elements.
• Existing theoretical models like Bethe's expression are energy-dependent and lack low-energy applicability.

Purpose of the Study:

• To predict electron stopping power (SP) for elements across the periodic table using machine learning (ML).
• To develop a model applicable for electron energies from 1 eV to 100 keV.
• To address the scarcity of experimental data and limitations of theoretical models.

Main Methods:

• Ensemble machine learning (ML) algorithms were employed.
• A small experimental database of electron SPs was used for training.
• Stacked generalization was utilized to enhance prediction accuracy.

Main Results:

• The ML model accurately predicted SPs for 54 elements, including 42 not in the training set.
• Predictions covered a wide energy range (1 eV to 100 keV).
• ML-predicted SPs showed excellent agreement with experimental data and outperformed other theoretical approaches.

Conclusions:

• Ensemble ML, particularly stacked generalization, provides accurate electron SP predictions.
• The ML model bypasses the need for complex physical parameters like dielectric functions.
• This approach enables SP prediction for elements lacking theoretical data.