Description
With neutrino oscillation physics firmly established, neutrino experiments have entered the era of precision measurement with limited budget for systematic uncertainties. One significant source of systematics stems from mis-modeling of detector physics processes, for which accurate and precise detector calibration is critical. A traditional approach in a neutrino experiment calibrates individual physics processes one by one, assuming little dependency among each other and missing the opportunity to minimize discrepancies between data and simulation through joint optimization of all models simultaneously. Furthermore, the analysis requires a separate “calibration” software and the manual effort spans typically a few years after physics data taking starts. These challenges can be addressed by an innovation of differentiable physics modeling. Taking an advantage of powerful gradient-based optimization algorithms which powered the explosive progress of deep learning recently, a differentiable simulator can solve the inverse problems including model parameter (self-)calibration as well as input data reconstruction (i.e. unfolding of the detector effects) directly using real data. Concretely, our application demonstrates the automation of simultaneous optimization of all physics models involved by directly minimizing the discrepancy metrics between data and simulation. This eradicates the need of having a separate software as the simulator can calibrate itself, and reduces both human effort and the time required to completely calibrate the whole detector. In addition, backed by well established analytical physics models, the simulator provide a high quality gradient information and enables a joint optimization of other diferentiable tools such as AI/ML-based data reconstruction and/or differentiable event generators. In this poster, we present the demonstration of a differentiable Liquid Argon Time Projection Chambers and Water Cherenkov detectors including two concrete applications: detector calibration and physics reconstruction through gradient-based optimization.