A Differentiable Neural Surrogate for Photon Propagation in Neutrino Telescopes

The researchers introduced a differentiable SIREN neural network called Candela, designed specifically for simulating the propagation of billions of photons in the IceCube neutrino telescope through ice or water. This model is based on Monte Carlo training data, generating complete events by decomposing the energy deposition of charged particles into point sources and superimposing the predicted responses. Compared with existing methods, Candela’s generation speed is 50 to 100 times faster, and its computational cost increases only slightly with neutrino energy. In terms of accuracy, the median photon yield deviation from Monte Carlo expectations is kept within 2%, and the time distribution accuracy reaches statistical limits. Additionally, this model provides end-to-end gradients, offering a new approach for optimizing the properties of the scattering medium to reduce systematic errors.