Speaker
Description
Accurate modeling of optical parameters in Water Cherenkov detector including absorption length, scattering length, and PMT quantum efficiency, is essential for reliable event reconstruction, yet their simultaneous calibration remains challenging due to strong inter-parameter correlations and the absence of tractable analytical likelihoods.
We propose extending the Simulation-Based Inference (SBI) framework with Neural Likelihood Estimation (NLE), recently demonstrated for liquid scintillator detectors, to Water Cherenkov detectors. A Normalizing Flow Density Estimator (NFDE) is trained on Geant4-based Water Cherenkov simulations to learn the conditional likelihood p(x | φ), where x comprises geometric summary statistics derived from reconstructed Cherenkov ring topology, and φ denotes the optical detector parameters.
The multi-dimensional ring-based summary statistics enable simultaneous constraint of correlated optical parameters whose degeneracy cannot be broken by scalar observables alone. Combined with Bayesian nested sampling, the learned likelihood provides posterior distributions over detector parameters with uncertainties limited by statistics and near-zero systematic bias. This work represents the first application of NLE-based MC tuning to Water Cherenkov detectors, establishing a foundation for systematic studies of how optical parameter uncertainties propagate to physics observables.
