Speaker
Description
Next-generation neutrino experiments such as DUNE and Hyper-Kamiokande will operate massive, monolithic detectors at unprecedented beam intensities and event rates. This places enormous demands on simulation, both in the sheer statistical volume required by modern analyses and in the precision of detector modeling needed for systematic control in oscillation measurements. Traditional calibration methods struggle with high-dimensional parameter spaces, particularly when parameters are strongly correlated and difficult to disentangle.
Differentiable programming offers a path forward. Built on JAX's automatic differentiation, we construct detector simulations in which every parameter is directly optimizable. Calibration can then scale to thousands of parameters, including spatial and temporal variations in detector response that are inaccessible to conventional methods. The same framework unifies simulation, reconstruction, and calibration, replacing tasks that have traditionally required separate pipelines.
We present open-source, GPU-parallelized, differentiable detector simulation frameworks for both Liquid Argon Time Projection Chambers and Water Cherenkov detectors. Both employ a modular design in which neural network surrogates replace components that lack precise analytical models, while the rest of the simulation chain remains fully interpretable. For LArTPCs, we implement a full detector simulation with a neural surrogate for optical photon transport, reducing per-event simulation time from roughly 50 seconds to sub-second. For Water Cherenkov detectors, we implement differentiable photon propagation with grid-based spatial acceleration, reducing simulation from minutes to milliseconds.
End-to-end differentiability transforms how both reconstruction and calibration are performed. Automatic differentiation provides the exact gradient of the likelihood with respect to every parameter in the simulation simultaneously, giving the steepest direction of improvement at each step. For reconstruction, this enables a direct fit of particle energy, position, and direction against observed detector response, achieving comparable performance to state-of-the-art reconstruction methods. For calibration, all detector parameters can be fitted simultaneously, disentangling strong correlations that are intractable for traditional methods.
These speeds further enable the release of large-scale open datasets for both detector types for machine learning research and development. From scalable calibration of evolving detector conditions to end-to-end optimization across the full analysis chain, differentiable detector simulation provides new tools for meeting the demands of next-generation neutrino experiments.