HELLMM: Heterogeneous Electrocatalysis with LLM-driven Mechanism Elucidation

The mechanism screened in a computational electrocatalysis study is still usually written by hand, and assumed to consist only of proton-coupled electron transfers. We present HELLMM, a pipeline in which a language model selects transformation rules, enumerates surface intermediates and filters them for chemical plausibility, while deterministic graph constraints and non-LLM atomistic and computational-hydrogen-electrode calculations evaluate what survives. Repeating it 23 times across three model checkpoints and four catalyst systems separates two sources of error. Conditional on a mechanism, the computed quantities reproduce to 0.5 mV; the run-to-run spread comes instead from the model selecting between competing mechanisms that are individually valid, partitioning exactly four and four on Pt(111) ORR and five and five on CO₂RR. The same energetics miss the literature Pt(111) ORR limiting potential by 0.5 V, an order of magnitude more than mechanism selection moves it and three more than the within-mechanism spread, so reproducibility and accuracy are separable problems. The audit also exposes failures that aggregate metrics hide: a descriptor blind to costly chemical steps, and prompt examples that bias enumeration.