Description
Precise and accurate Modeling of detector physics processes are keys to minimize the systematic uncertainties for experimental measurements. While detector physics model parameters can be optimized through the detector calibration analyses, the result is only as good as how the model can represent the real physics processes. Missing and incorrect analytical models cannot be addressed by detector calibration. Moreover, calibration of a Liquid Argon Time Projection Chamber, the detector choice of the short and long baseline neutrino experiments in the U.S., takes significant human and computational resources for software development, maintenance, and careful applications over the scale of years. In this project, we have addressed these challenges using SIREN, an implicit neural representation technique. SIREN, as a simulation software, can be optimized by directly minimizing the discrepancies between data and simulation without the need of developing a separate calibration software. As a neural network, SIREN is a universal function approximator and can represent any underlying physics models in real data. By design, unlike vanilla neural networks, SIREN guarantees the derivative and integral of the target physics model accurately. The last point implies that SIREN can be used to estimate the model gradient accurately, and hence it can be chained with other differentiable algorithms to jointly optimize for a broad set of applications, or propagation as well as estimation of model uncertainties. In this poster, we will present our application of SIREN for modeling physics of optical signals in LArTPCs. Furthermore, we will demonstrate its strength of differentiability through two applications including data reconstruction and data-driven optimization of other differentiable tools through joint-optimization techniques.