Predicting electronic stopping powers using stacking ensemble machine learning method

TL;DR

This study develops a stacking ensemble machine learning model that accurately predicts electronic stopping powers for various ion-target combinations across a broad energy spectrum, aiding applications in radiation physics.

Contribution

The paper introduces a novel stacking ensemble machine learning approach combining five models to predict electronic stopping powers with high accuracy over a wide energy range.

Findings

Model achieved R2=0.9985 on training data.

Model achieved R2=0.9955 on test data.

Predictions are highly accurate across diverse ion-target pairs.

Abstract

Purpose: Accurate electronic stopping power data is crucial for calculating radiation-induced effects in various applications, from dosimetry and radiotherapy to particle physics. In this study, Stacking Ensemble Machine Learning (EML) algorithm was developed to predict electronic stopping power for any incident ion and target combination over a wide range of ion energies. For this purpose, five ML models, namely BR, XGB, AdB, GB, and RF, were selected as base and meta learners to construct the final Stacking EML. Methods: 40,044 experimental measurements, from 1928 to the present, available on the International Atomic Energy Agency (IAEA) website were used to train machine learning (ML) algorithms. This database consists of 593 ion-target combinations across the energy range of 0.037 to 985 MeV. For model training, the eleven most important features were selected. The model evaluation was performed using several error metrics, including R-squared (R2), root-mean-squared-error (RMSE), mean-absolute-error (MAE), and mean-absolute-percentage-error (MAPE), on both the training and test datasets. Results: Based on model performance evaluation tests, a stack of eXtreme Gradient Boosting (XGB) and Random Forest (RF) via Bagging Regressor (BR) meta-learner had the highest lowest error margin. The value of R2=0.9985 indicated a near-ideal fit to all samples in the training data across the entire range of stopping powers. R2=0.9955 for predictions made by the model on the unseen test data suggested that the model accurately predicted the test data. Conclusions: The developed model resulted in highly accurate predictions for any ion-target combination across the whole particle energy spectrum. The associated model can serve as a universal tool to generate the stopping power data in a wide range of cases, regardless of the availability of experimental data or reliable theoretical equations.