High-performance statistical methods for reactor neutrino oscillations

We present a PyTorch-based framework for forward folded reactor neutrino spectrum fitting that accelerates the two main bottlenecks: IBD mapping and detector response, using (i) result caching, (ii) banded sparse matrices, and (iii) blocked construction of the response. On an Intel Xeon Gold 6338 CPU, these techniques reduce per-fit walltime by ≈ 7× ≈ 7 × (median over 5 runs) relative to a dense, unoptimized implementation, with <10⁻⁶ < 10 - 6 relative spectral error versus a double-precision baseline. The framework has been applied to reactor-neutrino oscillation analyses and is reusable in other neutrino experiments that rely on forward-folded energy spectra, enabling practical Feldman–Cousins coverage studies and large parameter scans at substantially lower computational cost.