When Can a Molecular-Dynamics Simulation Stop?

Molecular-dynamics (MD) simulations are usually run for a prespecified duration, even when the goal is to estimate an equilibrium observable to a stated precision. Stopping as soon as an estimated uncertainty becomes small enough is attractive, but repeated inspection invalidates uncertainty statements calibrated for a fixed time. We formulate data-dependent stopping as an inferential problem with an explicit precision target and a simultaneous-in-time coverage requirement, and organise it into four gated stages: temporal prediction against strong baselines, uncertainty calibration, an anytime-valid stopping rule, and measured high-performancecomputing (HPC) savings. We report exploratory first-stage results on four incrementally assembled ATLAS cohorts, using sequence-grouped cross-validation and a target computed only from frames after a 20 ns prefix. In the largest cohort (757 systems, 1,243 replicas, 689 sequence groups), ridge regression on 48 prefix summaries and a residual gated recurrent unit (GRU) both reduced RMSD mean absolute error relative to last-value persistence (0.610 and 0.636 versus 0.673 Å; paired bootstrap intervals excluding zero), whereas the GRU–ridge difference was not significant in any cohort. A genuine structure-only baseline, a compatible mdCATH analysis, sequential coverage and HPC cost measurements are still outstanding, so no stopping policy or computational saving is claimed.