AI and the World's Hard Problems · Chemistry & materials · arXiv · 2026-10-21
LatticeMind: A Validation-Gated Agentic Framework for Density Functional Theory Workflows
First-principles calculations are multistage workflows, and language models that can write a valid single input lose reliability once stages are chained, as dropped dependencies, wrong file handoffs or hallucinated keywords can invalidate results even when the solver runs. LatticeMind converts natural-language requests into executable, checkable DFT projects: a workflow contract specifies required stages, files and dependencies, deterministic validators inspect the project before any solver launch, and failed checks trigger bounded repair. On a 44-case RESCU benchmark, three non-flagship LLM backends reproduced every reference value, with 119 of 132 runs passing on the first validated plan; the same architecture reproduced every reference value on a 19-case VASP benchmark and generated and validated the 23 inputs of a multistage study of a Rh₆/InₓGa₁₋ₓN photocatalyst interface.