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

Machine-learned interatomic potentials simulate molecules at a small fraction of the cost of quantum chemistry, but they are only useful if they can tell us when they are guessing. The standard way to ask is to train a committee of predictors and measure how much they disagree. A multi-head committee makes this almost free for energies: M prediction heads share one network body, so a single forward pass through the network returns all M energy estimates at once. Forces are not free. A force is the derivative of the energy with respect to the atomic positions, and recovering each head's forces takes its own backward pass through that network - so force disagreement, the quantity that actually decides which structures a simulation should stop and verify, costs roughly M times more than energy disagreement. We ask whether a handful of cheap random probes can estimate that disagreement without computing all M sets of forces, using a pretrained eight-head committee on four molecular systems. The answer is no for replacement and yes for screening. Even given six of the seven probes the exact computation would use, the best estimator identifies only 72-80% of the genuinely most uncertain structures, so an approximate score cannot stand in for the exact one. But a calibrated filter built from that same score safely skips 84-87% of the expensive computations while still catching 96-98% of the uncertain structures, and beats cheaper alternatives in 11 of 12 comparisons. Whether this saves wall-clock time turns out to depend on how the model is run: filtering is worth 1.85x when structures are processed in batches, but only 1.11x one at a time, where it becomes a net slowdown. The decision-quality result is robust; the speed result is not.