Machine Learning Flags Fast-Neutrino Flavor Instabilities

Core-collapse supernovae and neutron-star mergers are among the densest, hottest sites in the cosmos. Elusive neutrinos carry off most of the released energy, and their ability to swap "flavors" can run away into fast instabilities that reshape both the explosion and the newborn neutron star. I trained a heavily regularized ∼4-6-layer, 100s of neurons-wide multi-layered neural network (MLNN) in PyTorch on 8 × 10^5 labeled points spanning seeded-unstable, guaranteed-stable, and neutron-star-merger snapshots; dropout, weight decay, batch normalization, early stopping, and an extensive hyper-parameter search kept it honest, and it now flags flavor conversion instability with F1 ≈ 0.95 in very little time. By slotting this network directly into radiation-hydrodynamics codes, I turn what was previously an impossible grid-by-grid dispersion solve into trivial overhead, letting full simulations evolve flavor physics alongside matter in real time.