·
PHIN-GAN — Physics-Informed Generative Surrogate
A Wasserstein GAN for particle transport with a second loss derived from physics. The analytical energy-loss distribution can be evaluated anywhere in phase space, so the generator is held to it even where the training data runs out.
Key result
~100×
faster than Geant4 on GPU
For large batches on a GPU. The generated steps are statistically indistinguishable from Geant4 (energy-distance p-values 0.76, 0.75 and 0.94 for the three generators against a Geant4-vs-Geant4 reference). The same GAN without the physics loss gives p-values near 0.
Resources