Force sketching in MD: Can we replace multi-head committees with screening methods for faster force uncertainty compute?

Machine-learned interatomic potentials simulate molecules far more cheaply than quantum chemistry, but a simulator-based workflow only benefits if the emulator can say where it is misspecified — that is what decides when to fall back on the expensive mechanistic model. Committee disagreement is the usual signal, and a multi-head committee makes it nearly free for energies: M heads share one trunk, so one forward pass gives all M energies. Forces are not free. Forces are gradients, and each head's forces need their own backward pass, so force disagreement — the quantity acquisition actually uses — costs about M times more. We ask whether a handful of cheap random probes can estimate it instead, on a pretrained eight-head committee across four molecular systems. The answer is no for replacement and yes for screening. Even with six of the seven probes the exact computation would use, the best estimator recovers only 72–80% of the genuinely most uncertain structures. But a calibrated filter built from the same score safely skips 84–87% of the expensive computations while still catching 96–98%, beating cheaper alternatives in 11 of 12 comparisons. The wall-clock benefit is regime-dependent: 1.85× in batch, only 1.11× one structure at a time. The decision-quality result is robust; the speed result is not.