AI summaryⓘ
The authors tested using machine learning models to speed up a very slow part of a complicated physics simulation involving neutrino radiation and matter interactions. They found three main problems: the slowest part only takes up about 17% of the total time, so even perfect improvements can't speed things up much; measuring model accuracy before using them doesn’t predict how well they perform in the real simulation; and the system often encounters situations very different from the training data, causing models to defer almost all work and slow things down. They also observed that errors accumulate steadily over time rather than randomly. Overall, their results show that learning surrogates face fundamental challenges in this context.
multiphysics simulationsurrogate modelsgeneral-relativistic radiation-hydrodynamicsneutrino radiationNewton solverAmdahl's lawout-of-distribution detectionmodel fidelityautoregressive error accumulation
Abstract
Learned surrogates for expensive inner solver blocks are a widely pursued route to faster multiphysics simulation. We report a controlled, end-to-end negative result and identify three structural barriers, none of them a deficiency of the network we trained. The testbed is the implicit Newton solve coupling energy-dependent neutrino radiation to matter in a general-relativistic radiation-hydrodynamics code, its most expensive physics routine per call. First, per-call cost and share of runtime are different quantities, and only the second bounds acceleration. An exclusive self-time profile puts the target block at 16.9% of critical-rank wall clock, capping any surrogate at ~1.2x by Amdahl's law. A surrogate 5.8x cheaper per call merely ties the solver, and the configuration stable enough to run without fallback reaches only parity. Second, offline accuracy cannot rank surrogates for deployment: across fourteen networks the pooled Spearman error-versus-survival correlation (rho = +0.73) is a between-family confound that vanishes under control (rho = -0.04). Third, a correct out-of-distribution gate cannot accelerate a loop that leaves its training distribution. We give the break-even deferral fraction in closed form: because the visited states sit 73x off the data manifold, the gate defers 96.8 to 99.7% of cells, almost invariant to surrogate quality. Including its own cost, the gated loop is a 0.94 to 0.96x slowdown. We further separate stability from fidelity: a never-crashing gated run accumulates a linear -19.9% density bias over 6000 steps. The error is a directed, ballistically accumulating bias, not the variance-driven divergence the autoregressive literature targets.