Low uncertainty High uncertainty Just queried (active)

Active Learning for Machine-Learned Interatomic Potentials (2D)

A 2D toy configuration space stands in for the atomic-environment space of a real machine-learned interatomic potential: two structural parameters (x, y) map to a synthetic but genuinely multi-featured "energy" surface. A real ensemble of Nadaraya–Watson kernel regressors, each with a different bandwidth, is trained on a growing labeled set and disagrees most where it has seen the least data. Every round, the engine reads that disagreement off the whole grid and queries ground truth at the single highest-uncertainty point — while an independent random-sampling baseline runs in parallel. The live learning curve plots prediction error against number of labels for both strategies, so the efficiency gain of active selection over random sampling is a measured result, not a claim.