Exhibit Hall | Forum 7
Purpose: Accurate electronic stopping power data is crucial for beam dosimetry in applications such as radiotherapy and particle research. The Bethe theory can be used to calculate stopping power of high energy incident ions, but it fails at lower energies. The majority of known experimental data, on the other hand, is only available for elements, limiting the validity of fitting formulas for more complex material compositions. We developed and optimized a machine learning (ML) model that can predict mass stopping power for any incident ion and target over a wide range of ion energies.
Methods: 40,044 experimental measurements were used to train several machine learning (ML) algorithms. For model training, the eleven most important features were taken into account. This model was evaluated using several error metrics including R-squared, root-mean-squared-error (RMSE), mean-absolute-error (MAE) and mean-absolute-percentage-error (MAPE), on train and test datasets for individual ion-target combinations. Finally, we assessed potential overfitting to ensure the predictive quality of the model.
Results: A stack of Boosting Regressor (BR) and Gradient Boosting Regressor (GBR) via Random Forest (RF) meta-model had the highest efficiency based on model performance evaluation tests. The calculated value of R2=0.9981 indicates a near-ideal fit to all samples in the training data across the entire range of stopping powers. R2=0.9942 for predictions made by model on the unseen test data suggests that the model accurately predicts the test data.
Conclusion: The developed model is capable of resolving problems such as missing experimental data and fitting parameter errors. This model can be used to calculate stopping power for a variety of elements, and compounds in combination with any incident ions across the whole particle energy spectrum. These capabilities of the presented model prove that machine learning approaches are well suited for our objectives.