LatticeMind preprint reports more reliable AI workflows for materials calculations

Researchers report that LatticeMind, an AI framework for coordinating calculations of material properties, reproduced every reference value across a 44-case RESCU benchmark using three language model backends. Of 132 runs, 119 passed on the first validated plan; the remainder passed after one repair cycle. The framework turns natural-language requests into density functional theory projects, which calculate electronic properties. It specifies required stages, files and dependencies, then checks them before launching the solver and repairs failed checks. A conventional trial-and-repair baseline achieved 71% success at twice the token cost per success. The framework also reproduced every reference value on a 19-case VASP benchmark, though the findings remain preprint results on the tested workflows.