Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation

Optimizing the Optimizer. Language Models Discover Faster Molecular Relaxation. By Artem Tsypin and six others, from AXXX, ITMO University, and one other institution. Published on the fifth of October. Its code repository is public and real: about twenty-five source files.

Geometry optimization is a major cost in many quantum-chemical workflows. Every optimization step needs one force evaluation, and at the density-functional level that evaluation dominates the wall time. Count the force calls and you have, near enough, counted the bill. Research in this area has already produced a broad range of optimization methods, so the question is a pointed one: can a language model improve on the best of them, through autoresearch?

The move is to put the optimizer itself up for revision. An agent rewrites the optimizer's code with a single objective, fewer force calls, and two admission gates restrain what it is allowed to keep. One gate rejects premature stopping, a saving that comes from quitting early rather than converging faster. The other rejects improvements that do not generalize to unseen molecules.

The search starts from Sella, described as the fastest open-source optimizer available. What comes out is AutoSella, a family of two optimizers. One detail shapes how the results should be heard: the agent used no DFT gradients during the search, so it never touched the expensive level of theory on which the headline is measured.

Both AutoSella variants deliver consistent force-call reductions relative to Sella, on held-out molecular benchmarks and on potentials that were not used during the search. The most notable result is at the r2SCAN-3c DFT level. There the best variant needs only forty point two to seventy-seven point two percent of Sella's force calls, while achieving the same energy reduction.

Put plainly, a language model started from the fastest open-source baseline and wrote an optimizer that reaches the same energy with fewer of the most expensive steps. The gates are what make that believable: they keep the agent from gaming its own metric, and the held-out tests show the savings travel.