Generative Adversarial Networks (GAN) for compact beam source modelling in Monte Carlo simulations
This is an incremental improvement for researchers in medical physics and computational simulations, enabling more efficient storage and generation of beam source models.
The authors tackled the problem of modeling large phase space files in Monte Carlo simulations by using a Generative Adversarial Network (GAN) to create a compact generator, reducing file size from gigabytes to 10 MB while achieving less than 1% voxel-by-voxel relative difference in deposited energy distributions compared to reference data.
A method is proposed and evaluated to model large and inconvenient phase space files used in Monte Carlo simulations by a compact Generative Adversarial Network (GAN). The GAN is trained based on a phase space dataset to create a neural network, called Generator (G), allowing G to mimic the multidimensional data distribution of the phase space. At the end of the training process, G is stored with about 0.5 million weights, around 10 MB, instead of few GB of the initial file. Particles are then generated with G to replace the phase space dataset. This concept is applied to beam models from linear accelerators (linacs) and from brachytherapy seed models. Simulations using particles from the reference phase space on one hand and those generated by the GAN on the other hand were compared. 3D distributions of deposited energy obtained from source distributions generated by the GAN were close to the reference ones, with less than 1% of voxel-by-voxel relative difference. Sharp parts such as the brachytherapy emission lines in the energy spectra were not perfectly modeled by the GAN. Detailed statistical properties and limitations of the GAN-generated particles still require further investigation, but the proposed exploratory approach is already promising and paves the way for a wide range of applications.