Deep Learning Guided Optimization of Primary Electron Source Parameters in Geant4 for Clinical Linac Modeling

Monte Carlo (MC) simulation is widely used in medical physics for radiation transport modeling, beam commissioning, and dose calculation. However, the optimization of primary electron source parameters for clinical linear accelerators using MC simulations can require many repeated simulations, making the process time-consuming.

This study aimed to develop a data-driven surrogate framework to guide the optimization of primary electron beam parameters in Geant4 for an Elekta Synergy linear accelerator operated in 10 MeV electron mode. The objective was to predict the gamma pass rate for the 2%/2 mm criterion and to identify the most influential beam parameters.

A Geant4-based Monte Carlo model was used to simulate the 10 MeV electron beam. Simulated dose distributions were compared with experimental measurements in water under reference conditions. A feedforward neural network was trained to predict the gamma pass rate from four primary source parameters: mean energy, energy spread, spatial spread, and angular spread. A Gradient Boosting Regressor was also used to evaluate the relative importance of these parameters.

The feedforward neural network showed high predictive performance, with an R² value of 0.9924 for the training dataset and 0.9857 for the testing dataset. The model enabled rapid screening of beam parameter configurations. The Gradient Boosting Regressor indicated that the mean energy was the dominant parameter influencing agreement between simulated and measured dose distributions.

The proposed deep learning-guided framework can reduce trial-and-error in Monte Carlo-based electron beam model tuning while maintaining clinically relevant accuracy. This approach may support more efficient optimization of primary electron source parameters for clinical linac modeling.