Kolmo

AI surrogates can simulate physics fast, but they fail silently: nobody tells you when the prediction stops being true. Kolmo pairs exact JAX spectral solvers with neural surrogates and measures trust explicitly. The solvers cover 2D Navier–Stokes, heat convection, β-plane cyclones, a PDE compiler (Cahn–Hilliard, Gray–Scott and more) and 3D turbulence. The surrogates are a hybrid neural cellular automaton (NCA), an FNO, a CNN and a JEPA-style latent world model. Kolmo measures error vs horizon, steps to failure, spectra and energy drift. A hybrid mode lets the solver step in whenever the AI drifts.