acceptodds
Under review as a conference paper at ICLR 2027

Learning Transition-Metal Complexes Reaction Barriers with Active Learning and Machine Learning Interatomic Potential

Abstract

Fine-tuning a machine learning interatomic potential (MLIP) to accurately model a chemical reaction needs training data along the reaction path. However, that path is exactly what is unknown beforehand. Here we demonstrate a method to efficiently build such a dataset from the only knowledge of the reactant and product. We show that fine-tuning only works when this method is robust. We use a combination of active learning and a custom path generator. This generator is named curved reactive scan, in which each reactive bond follows its own power law, feeding a certified saddle search whose by-products constitute the training set. As a result these data points sit on the reaction path. We show that the fine-tuning is data efficient: for 25 transition-metal complex reactions from the benchmark MOBH35 (15–67 atoms), we generate 86 to 284 density functional theory (ωB97M-V/def2- TZVPD) labels per reaction, and the fine-tuned versions of UMA S 1.2 improve the retrieving of their transition states, forward and reverse energy barriers, at chemical accuracy on average. Those results outperform the pretrained model, even though it was trained on OMol25, a data set that contains reactive configurations generated from these very MOBH35 systems. Our work shows that blind generation of molecular geometries along the reaction path is a crucial step towards efficient MLIP fine-tuning, to potentially unlock applications in metalloprotein studies, and transition metal catalyst discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning interatomic potential, active learning, chemistry, chemical reactions, molecules

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TDcROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This contribution presents an approach for determining reaction paths and the associated energetics for transition metal complex reactions, making use of machine learning interatomic potentials. The work is very clearly presented and makes use of the MOBH35 dataset. The presented future work is comprehensive. The key contributions are the path generator, the active learning procedure for the MLIPs across the pathway, and their results on 25 reactions. There is a good degree of open-sourcing of data/code planned.

Critical strengths

The approach is generalisable for transition state and reactive path searching across other chemical systems and is demonstrated to work well. The paper is clear in its contributions, limitations and targets for future work.

Critical weaknesses

Given the paper location, there is not significant development of machine learning, the focus is more computational chemistry with machine learning elements as a tool, but this is a very minor criticism.

Questions for authors

None

Reviewer_NwyNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper describes an active learning workflow for fine-tuning pre-trained MLIPs for a specific class of molecules, using a newly introduced algorithm for transition path search.

Critical strengths

The new transition state search algorithm seems interesting (but potentially not for the ICLR audience).

Critical weaknesses

Lack of ML novelty. No new ML method, model, or approach is introduced. The contributions are a transition state path generator (not in the scope of ICLR), an automated active learning pipeline for MLIPs (lack of ML novelty), two domain-specific metrics (not highly relevant to MLIPs beyond the specific task here), evaluation results on a specific dataset studied here (without any benchmarking), and a comparison of a related method for the first contribution (again, not in the scope of ICLR).

This weakness should be interpreted in the context of the ICLR venue, where novel machine learning methods and approaches are published. In the context of computational chemistry, where those approaches are used to solve chemistry challenges, the weaknesses mentioned above are invalid, which is why I encourage the authors to submit the manuscript elsewhere.

Questions for authors

What is your hypothesis that your scan method leads to better results than NEB? What was the motivation not to just use NEB in the first place?

Reviewer_VHkBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts a pretrained interatomic potential to individual transition-metal reactions through active learning. A curved scan of reactive bond distances supplies candidates for saddle optimization, filtering, and DFT labeling. The resulting models improve barrier estimates with relatively little reaction-specific data. A smaller fixed-budget comparison also examines scan-generated data against an NEB-based alternative.

Critical strengths
  1. Generating useful training structures is a real difficulty for these reactions. The curved scan is a practical way to explore asynchronous bond changes without searching a full grid of bond distances.
  2. The evaluation distinguishes errors at reference structures from errors at structures found by the model. This matters because fitting barrier energies does not necessarily mean that the potential can find the desired saddle. The fixed-budget comparison in Section 5.2 also keeps downstream selection and training consistent.
Critical weaknesses
  1. The stopping rule uses reference barrier errors (Section 4.2), while the proposed workflow is described as starting from endpoints. Those reference errors would not be available for a new reaction with an unknown transition state. Section 5.1 also refers to a best-round checkpoint without clearly stating how it is selected. This makes the reported label efficiency difficult to interpret. The smaller fixed-budget comparison helps, but does not resolve the issue for the main campaign.
  2. The advantage over NEB is not firmly established by the main comparison. Section 5.2 reports a lower error for the scan on every tested reaction, but most differences are small and comparable to the replicate spread reported for the one reaction with repeated runs. The clearest separation comes from a single case. The main campaign also uses only one seed per reaction. The results suggest that the scan can be useful, but support a narrower claim than a general improvement in robustness and data efficiency. Comparing the existing learning curves at smaller label budgets and reporting path-generation times would help assess the efficiency claim.

UMA was pretrained on configurations derived from MOBH35, as the paper acknowledges. The results therefore demonstrate adaptation on systems represented in pretraining; they do not establish the same benefit for unfamiliar systems. This does not make the improvements unimportant, but the scope should be clear.

The reactive-bond detection rule also needs a precise description. Section 3 says pairs are close enough to form a covalent bond and change separation between endpoints, but does not give the distance criteria. This matters for ambiguous metal–ligand bonds, especially since the paper itself suggests that a missed reactive coordinate may explain a failure.

The Hessian and connectivity checks operate on the learned potential. Please distinguish those checks from DFT confirmation of a saddle, and explain the omitted MOBH35 reactions. The related-work discussion should also include earlier active surrogate-assisted searches, such as GP-NEB and AL-NEB. The NEB comparison does not isolate what the curved schedule adds over an otherwise matched linear scan.

I lean toward rejection because reference-independent label efficiency remains unclear and the evidence for a general advantage over NEB is limited.

Questions for authors
  1. What determines termination and checkpoint selection in the main campaign? Can the saved runs be compared at a fixed label budget without using reference transition-state information?
  2. Do the existing learning curves show a clearer scan advantage at smaller label budgets? What is the path-generation time for each method, and how much of the observed difference is supported by repeated runs?
  3. What distance and distance-change criteria define a reactive pair, particularly for metal–ligand bonds? How sensitive is the detected set to these choices?
  4. Why were the omitted reactions excluded, and which model-found structures have been checked by DFT refinement, frequencies, and endpoint connectivity?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.