Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation

Geometry optimization is a principal cost in many quantum-chemical workflows: each optimization step requires one force evaluation, and at the density-functional level that evaluation dominates the wall time. Decades of analytical work have produced a set of optimizers, and we ask whether a language model can improve on the best of them with autoresearch. An agent rewrites the optimizer itself to minimize force-call counts, restrained by two admission gates that reject premature stopping and improvements that do not generalize to unseen molecules. Starting from Sella, the fastest open-source optimizer available, the search produces AutoSella, a family of three optimizers. All of them deliver consistent force-call reductions relative to Sella across held-out molecular benchmarks and potentials not used during the search. Most notably, at the \texttt{r2SCAN-3c} DFT level, the best variant requires only $73.5$-$89.7\%$ of Sella's force calls while achieving essentially the same energy reduction, even though the agent used no DFT gradients.