Language Models Rewrite a Molecular Optimizer to Cut Force Evaluations

Introduction

Molecular geometry optimization is a routine but expensive part of quantum-chemical workflows. The structure is updated step by step, and each step generally requires a new energy and force calculation. At the density-functional theory level, these evaluations can dominate the total runtime. Reducing the number of force calls can therefore deliver a more direct speedup than making only low-level implementation changes.

A study titled “Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation” explores whether a language-model agent can improve the optimization algorithm itself. Rather than asking the model to predict molecular structures, the researchers allowed it to rewrite Sella, an open-source molecular geometry optimizer, with the objective of reaching comparable relaxation quality using fewer force evaluations.

Key points

• A strong starting point: The search began with Sella, described in the supplied material as the fastest open-source optimizer available for this task. The goal was to reduce its force-call count without giving up optimization quality.
• Cheap search, expensive deployment: Candidate changes were explored with inexpensive GFN2-xTB calculations. The agent did not use DFT gradients during the search, which kept experimentation cheaper and made the later transfer result more notable.
• Two admission gates: A candidate was rejected if it stopped too early or if its apparent gains failed to generalize to unseen molecules. These gates were designed to prevent speedups that merely reflect weaker convergence or overfitting to the search set.
• Two AutoSella variants: The process produced a family of two optimizers. Both reduced force calls relative to Sella on held-out molecular benchmarks and on potentials not used during the search.
• Transfer to DFT: On r2SCAN-3c DFT tests, the strongest variant required 40.2% to 77.2% of Sella’s force calls while achieving the same energy reduction. That corresponds to a reduction of roughly 22.8% to 59.8%.

Why it matters

The important idea is not simply that a language model generated faster code. It is that the model was placed inside a measurable, automated research loop: propose an algorithmic change, run experiments, check reliability and generalization, and retain only changes that pass predefined criteria. This setup is particularly suitable for numerical optimizers, where performance can be evaluated with clear resource and convergence metrics.

The cross-level result is also significant. Searching with a cheaper potential yet obtaining improvements on DFT suggests that some useful optimizer behaviors are not tied exclusively to the search environment. At the same time, the supplied material reports benchmark results rather than universal superiority. It does not establish that AutoSella wins for every molecule, potential, or convergence regime, so practical users should still check stability and final optimization quality in their own workflows.

More broadly, this work points toward language models contributing to scientific software beyond code completion. They may help explore update rules, stopping criteria, and resource-allocation strategies in established numerical algorithms. AutoSella offers a concrete template: let the agent generate candidates, but let reproducible experiments and held-out tests determine whether those candidates become part of the tool.

Source: Hugging Face Daily Papers