Speaker
Description
Liquid argon time projection chambers (LArTPCs) at DUNE and SBN need fast simulation, automated event reconstruction, and precise detector calibration, which are usually handled by separate tools. We built JAXTPC, a single differentiable detector simulator that unifies all of three tasks. Powered by JAX for efficient parallelization, JAXTPC models the full LArTPC response, from charge recombination and electron drift through the diffused field response, electronics, and noise, across detector geometries and wire or pixel readout. It runs about 100× faster on GPU than conventional CPU pipelines, making large, fully-labeled datasets practical to generate; a multi-million-event open dataset is in production.
The same model opens new ways of addressing reconstruction and calibration through automatic differentiation and high dimensional gradients. Reconstruction recovers an event from its wire signals, either as a parametric particle or as free charge segments initialized by a neural network. Calibration tunes detector parameters such as recombination, electron lifetime, and diffusion simultaneously. Optimization quality depends heavily on the choice of loss. An electrostatics inspired Sobolev loss matched to the sparse wire images enables both processes by providing a richer landscape. In closure, reconstructed charge agrees with truth to better than a percent, geometry to sub-millimeter, and the recovered calibration parameters to a few percent.
