Light Antinuclei Coalescence: Femtoscopic Constraints via Neural-Flow Surrogates

Precise predictions of cosmic-ray antinuclei fluxes, a prime dark matter signature, are limited by the lack of data constraining production rates of antinuclei. We mitigate this bottleneck with a fast, differentiable normalizing-flow surrogate for the femtoscopic source model (CECA), fit to 49 ALICE proton-proton correlation functions, and extrapolated via scaling laws to the low-multiplicity domain relevant for cosmic rays. The surrogate reproduces CECA with sub-percent emulation fidelity, yielding data-constrained source functions that remove the dominant uncertainty in coalescence-based antinuclei production rates. The resulting uncertainties on the coalescence parameters $B_2$ and $B_3$ shrink from factors of 10-100 and ${\sim}$1000 to the few percent and ten-percent level, respectively, with an additional ${\sim}15\%$ wavefunction systematic for $B_{2}$.