Generative adversarial networks (GAN) for compact beam source modelling in Monte Carlo simulations

Area of Science:

• Computational physics
• Medical physics
• Machine learning

Background:

• Monte Carlo simulations often require large phase space files, posing storage and computational challenges.
• Efficiently representing complex multidimensional data distributions is crucial for accurate simulations.

Purpose of the Study:

• To develop and evaluate a compact generative adversarial network (GAN) for modeling large phase space files.
• To reduce the data size of phase space files while preserving essential distribution characteristics for simulations.

Main Methods:

• A generative adversarial network (GAN) was trained on phase space datasets to create a Generator (G) neural network.
• The Generator (G) mimics the multidimensional data distribution of the original phase space.
• The GAN-generated particles were used in simulations and compared against reference datasets from linear accelerators and brachytherapy seed models.

Main Results:

• The trained GAN model (Generator) is stored compactly (approx. 10 MB) compared to original files (few GB).
• Simulations using GAN-generated particles showed 3D deposited energy distributions close to reference data (<1% voxel-by-voxel difference).
• Sharp spectral features, like brachytherapy emission lines, were not perfectly modeled, indicating areas for improvement.

Conclusions:

• The GAN-based method offers a promising, data-efficient approach to modeling phase space files for Monte Carlo simulations.
• This technique has potential applications in medical physics, particularly for linear accelerator and brachytherapy simulations.
• Further investigation into the statistical properties and limitations of GAN-generated particles is warranted.