Generalizable machine learning framework for Bragg peak prediction in polymeric materials using Monte Carlo simulations

Abstract

Accurate prediction of ion-induced energy deposition in polymeric materials is essential for space radiation shielding and advanced radiotherapy applications. This study proposes a generalizable machine learning framework for predicting Bragg peak characteristics using Monte Carlo simulation. Energy deposition profiles were generated via SRIM/TRIM and Geant4 for ions with atomic numbers Z = 1–10 at energies between 70 and 150 MeV. Two key targets were defined: Bragg peak position (MM 150 ) and maximum ionization intensity (IONIZ 150 ). For SRIM/TRIM data, the Poly-2 method achieved near-ideal accuracy for MM 150 (R 2 = 0.9999), while persistence performed best for IONIZ 150 (R 2 = 0.9991). However, performance degraded in Geant4 data, where MM 150 errors increased (for Poly-2) and IONIZ 150 showed higher variability (for sMAPE). The proposed machine learning framework, evaluated using LOPO validation, achieved consistently high generalization performance. In SRIM/TRIM, R 2 reached 0.9942 ± 0.0178 (MM 150 ) and 0.9951 ± 0.0144 (IONIZ 150 ), with sMAPE values of 5.11% and 3.76%, respectively. In Geant4, similar robustness was observed, with R 2 up to 0.9942 and sMAPE reduced to 4.41% for IONIZ 150 . These results demonstrate that simulation-informed machine learning enables accurate, robust, generalizable prediction of Bragg peak behavior in polymeric materials, offering a scalable framework for radiation interaction modeling and material design.