CrystalJev cuts materials screening cost 30x by reading foundation models as fast decision-makers

CatAstro_Piyush · x · 2026-10-08

The CrystalJev paper reframes atomistic foundation models as fast, calibrated decision-makers rather than slow simulators:

One forward pass of a frozen interatomic potential per unrelaxed structure answers typed stability questions with calibrated probabilities, finite-sample guarantees, and a rule for when to "think slowly".
Across 65 Matbench Discovery models, a "stable" call is shown to be a probability in disguise, explained by model errors and candidate population.
A single pass decides nearly as well as full relaxation at 1/30th the cost; value-of-information theory routes slow computation only where decisions can change.
The same layer answers electronic, mechanical, and molecular questions.
In a registered prospective test with 700 new DFT calculations, single-pass forecasts overstated the stable fraction (5.8%) of unseen candidates by at most 2.1 percentage points.

More from Research