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

Area of Science:

• Medical Physics
• Computational Physics
• Radiotherapy Physics

Background:

• Monte Carlo (MC) simulations are crucial in medical physics for radiation transport, beam commissioning, and dose calculations.
• Optimizing primary electron source parameters for clinical linear accelerators with MC simulations is often time-consuming due to repeated simulations.

Purpose of the Study:

• To develop a data-driven surrogate framework for optimizing Geant4 primary electron beam parameters for an Elekta Synergy linear accelerator (10 MeV electron mode).
• To predict the gamma pass rate (2%/2 mm criterion) and identify influential beam parameters.

Main Methods:

• A Geant4 MC model simulated the 10 MeV electron beam, with dose distributions validated against experimental measurements.
• A feedforward neural network was trained to predict gamma pass rates from four source parameters (mean energy, energy spread, spatial spread, angular spread).
• A Gradient Boosting Regressor assessed parameter importance.

Main Results:

• The neural network achieved high predictive performance (R²=0.9924 training, R²=0.9857 testing), enabling rapid screening of beam parameters.
• The Gradient Boosting Regressor identified mean energy as the most influential parameter for dose distribution agreement.

Conclusions:

• The deep learning framework accelerates MC-based electron beam model tuning while maintaining clinical accuracy.
• This approach can enhance the efficiency of optimizing primary electron source parameters for clinical linac modeling.