← News

ResearchSources: arXiv · 18 Nov 2026

Popcornn optimizes neural network paths to find transition states on machine learned potentials

The arXiv paper introduces a transition state search framework built on machine learned interatomic potentials, reporting more correct finds than nudged elastic band and growing string methods across thousands of organic reactions.

In plain words

Chemists need the transition state of a reaction — the fleeting midpoint structure that decides how fast a reaction runs and what it produces. The paper's Popcornn framework treats the search as a continuous neural network path on a machine learned potential, a fast trained surrogate for quantum chemistry calculations. The authors say this finds more correct transition states than standard double-ended search methods.

What was announced

Transition states control chemical reaction rates and outcomes, and locating them remains a central challenge in computational chemistry, according to the arXiv abstract. Double-ended search methods start from known reactant and product structures, but the paper says they generally rely on discretized path representations developed for density functional theory, where evaluating the potential energy surface is expensive. Machine learned interatomic potentials (MLIPs) cut that cost sharply and, through automatic differentiation, make higher-order derivatives cheap to obtain.

Popcornn instead optimizes continuous neural network paths on top of an MLIP using a parallelized adaptive numerical integrator. The authors report that this removes path discretization artifacts, reduces hyperparameter sensitivity, and allows non-local optimization that uses second-order potential energy surface information to get past local features. In their tests across thousands of organic reactions and multiple MLIPs, Popcornn consistently recovered more correct transition states than nudged elastic band and growing string methods. The framework is described as flexible in path objective, integration strategy, and optimization algorithm, and as suited to high-throughput search. These results come from the authors' own benchmarks and have not been independently confirmed.

What it does not show

  • The evidence is the paper's abstract only; benchmark details, reaction sets, and failure cases are not available in the supplied passages.
  • Reported advantages over nudged elastic band and growing string are the authors' own claims, not independently verified.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.