Differentiable surrogate for modeling the physics of optical propagation in a LArTPC

The liquid argon time projection chamber (LArTPC) is a common detector technology for accelerator-based neutrino experiments including the Deep Underground Neutrino Experiment (DUNE) and Short Baseline Neutrino (SBN) program. A typical LArTPC event produces 10^6 to 10^9 photons, and the modeling of the propagation process has a high computational cost. A popular alternative to tracking individual photons is the lookup table, which contains the average probability of observing a photon per detector at discrete points in the detector and is created via simulation. SIREN, an implicit neural representation using sinusoid activations, is a powerful tool for parametrizing a variety of signals as continuous functions. We have developed two ways to train SIREN to model photon transport in LArTPCs: one via a lookup table, and one via calibration datasets generated from simulation or real data. Because SIREN parametrizes signals as continuous functions, the gradients of the function can be learned. Also, the model has been shown to address the challenge of scaling for large detectors like DUNE. Here, we demonstrate the ability to train SIREN using calibration datasets of the DUNE Near Detector (ND) LAr prototype – the 2 x 2 demonstrator – which will start collecting data in Spring 2024.