Measured acceleration · Materials science · Computer science
CrystalJev tests single-pass materials screening with prospective DFT labels
Calibrated single-pass screening rivals relaxed-potential decisions and is assessed against newly computed DFT stability labels.
Summary
CrystalJev reports screening decisions at about one-thirtieth of relaxation’s forward-pass cost after substantial in-domain training. In a registered 700-calculation campaign, 23 of 397 labeled random test candidates are stable under the specified DFT hull criterion. A separate top-ranked shortlist yields 98 of 100 stable labels. These are computational labels, not synthesized materials.
AI role
Small decision heads read frozen atomistic-model states and estimate when slower relaxation or DFT computation can change a screening decision.
Narrative role
Adds prospective computational screening evidence and shows how targeted slow computation can reduce model-evaluation burden in materials search.
Caveat
Code/data await publication; no independent reproduction. Forward-pass cost is not an end-to-end speedup. DFT labels, correlated failures, calibration shift and registration deviations limit interpretation; synthesis was not tested.