Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation
An agent rewrote Sella into AutoSella, needing only 40.2–77.2% of Sella's force calls at r2SCAN-3c.
Geometry optimization is costly because each step requires a force evaluation that dominates DFT wall time. A language-model agent rewrote the Sella optimizer under gates that reject premature stopping and non-generalizing gains, producing the AutoSella family. Both variants cut force calls versus Sella on held-out molecules and potentials not used in the search. At r2SCAN-3c, the best variant needed only 40.2–77.2% of Sella's force calls for the same energy reduction, without using DFT gradients during search.
- Search started from Sella, described as the fastest open-source optimizer.
- Gates reject premature stopping and gains that fail on unseen molecules.
- Best variant used 40.2–77.2% of Sella force calls at r2SCAN-3c.
- Reductions held on held-out benchmarks and potentials unused in search.
From the original article
1 min read at huggingface.co
Geometry optimization is a major 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. Research in this area has produced a broad range of optimization methods, and we ask whether a language model can improve on the best of them through autoresearch.
By Artem Tsypin, Vladimir Deshchenya, Kuzma Khrabrov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.