Speakers
Description
Precision reactor neutrino experiments rely on forward-folded reconstructed energy spectrum fitting to measure oscillation parameters and determine the neutrino mass ordering. These analyses require repeated convolution of reactor fluxes, inverse beta decay (IBD) cross sections, and detector response models. When constructing frequentist confidence intervals using the Feldman–Cousins method, large ensembles of toy Monte Carlo simulations are needed, making spectrum evaluation the dominant computational bottleneck.
We present a PyTorch-based high-performance framework that significantly accelerates reactor neutrino spectrum fitting while maintaining numerical precision (<10⁻⁶ relative spectral deviation). Profiling shows that over 98% of the runtime is spent in spectrum construction, particularly in IBD mapping and detector response matrix calculation. To address this, we implement three complementary optimizations: multi-level caching of reusable quantities, banded sparse-matrix (CSR) storage of the differential cross-section matrix, and blocked (tiled) computation of the detector response matrix to improve memory locality and reduce costly error-function evaluations.
Benchmarks on an Intel Xeon Gold 6338 CPU demonstrate an overall ~7× reduction in per-fit wall time. This scalable framework enables practical large-scale Feldman–Cousins studies and extensive parameter scans, substantially lowering the computational barrier for precision reactor neutrino oscillation analyses. This work has been published on Eur. Phys. J. C 85,1420 (2025).