Differentiable Simulation for Calibration and Reconstruction of Optical Particle Detectors

Résumé fourni par la source

Optical particle detectors face complex calibration challenges as experiments scale, with high-dimensional parameter spaces and strong correlations making traditional Monte Carlo sampling computationally prohibitive. We introduce LUCiD, a differentiable ray-tracing framework for optical particle detectors that computes expected detector responses by propagating probability weights rather than sampling discrete paths. Straight-through estimation handles stochastic path selection while Gaussian relaxations enable differentiable hit detection, implemented in JAX for GPU acceleration. The resulting gradients enable direct navigation of correlated parameter spaces where sampling methods struggle, unlocking new gradient-based approaches to calibration, reconstruction, and physics-ML integration.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.