acceptodds
Under review as a conference paper at ICLR 2027

Reactive Force Error and Reaction-Path Completion in Pretrained Potentials

Abstract

We test local force response, reactive force accuracy and reaction-path completion in pretrained interatomic potentials. Across five broad-domain checkpoints, local force-response directions closely match their references, while forward barriers are compressed on 200 Transition1x reactions (slopes 0.660–0.857). On near-path configurations, median absolute-force cosines range from 0.311 to 0.605. Under the primary fixed-endpoint CI-NEB protocol, 3–7 of 20 paths converge per checkpoint. Local response, reactive forces and path completion therefore need separate evaluation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning interatomic potentials, reaction pathways, force evaluation, pretrained models, nudged elastic band

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qbhjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, local force response, reactive force accuracy, and reaction-path completion are tested in pretrained interatomic potentials. A matched estimate of reactive force error is reported in this manuscript. Moreover, separate tests of prediction and path completion are also presented.

Critical strengths

1, The reported studies of machine learning interatomic potentials in this paper are generally interesting.

2, The presented results of this paper may have promising applications.

Critical weaknesses

1, Despite my great effort, many parts of this paper is very hard to understand. A major point is that this paper presents a lot of unexplained terms in the field of interatomic potentials, such as "Lanspace 17", "SevenNetOmni", "Transition1x reactions", "PBE0 energy", "CCSD(T) energy calculations", and so on (there are too many). Such terms make this paper difficult to understand in the ML venue.

2, Many quantitative results are reported, but authors do not explain well what such results indicate in many places. For example, in the Abstract, authors write "while forward barriers are compressed on 200 Transition1x reactions (slopes 0.660--0.857)" without explaining what slopes 0.660--0.0857 stand for Transition1x. This makes the significance of this paper hard to be understood.

3, The major contributions of this manuscript, just as authors present in the Contributions part in the Introduction, are testing exising methods in the interatomic potentials fields and reporting the finding results. Therefore, the novelty and technical significance of this manuscript are fairly limited.

4, Another major issue of this manuscript is lack of comparison with other baseline methods. Authors do not compare their own tested results or their proposed methodological approaches with other baselines. Meanwhile, authors even do not present a detailed and insightful background with comparison of their own proposed methods in their "Related Work" Section. Whether authors' methods and approaches are meaningful in the context of existing research such as (Morrow et al., 2023), (Yu et al., 2024), (Schreiner et al., 2022) is unclear.

5, For the soundness of this manuscript, although the results may inspire some research or promote some futural applications, the direct practical impact or significance of this manuscript is not clear. Authors do not present the clear direct advancement in the interatomic potentials promoted by this manuscript. This makes the practical soundness of this manuscript somewhat vague.

Questions for authors

1, Could you please explain the professional interatomic potentials terms, such as "Lanspace 17", "SevenNetOmni", "Transition1x reactions", "PBE0 energy", "CCSD(T) energy calculations" and many others in the paper?

2, Could you please explain what the main reported quantitative results indicate in the manuscript?

3, Could you compare your proposed approach with existing related methods? Whether your methods and approaches are meaningful in the context of existing research such as (Morrow et al., 2023), (Yu et al., 2024), (Schreiner et al., 2022) is unclear.

4, Could you demonstrate the practical soundness of this paper? Although the results may inspire some research or promote some futural applications, the direct practical impact or significance of this manuscript is not clear.

5, Could you present a detailed and insightful background introduction at the intersection of machine learning and interatomic potentials? The lack of such background makes it hard to tell the significance of this manuscript.

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

Review of "Reactive Force Error and Reactive Path Completion in Pretrained Potentials" In the presented manuscript, authors highlight the fact that several of these large MLIPs provide accurate single point computations, but it is not the complete picture as often, the reaction pathways contain significant errors. The presented manuscript extensively verifies it against DFT and CC methods. The use of CCSD(T) is highly commendable, as often the reactive pathways using just DFT are insufficiently resolved, and are method dependent. The comparisons are sound, and methodology is also reasonable.

Critical strengths

The paper has enumerated reasonable mix of MLIP, DFT methods, and higher level computations. It has identified a key part of the problem in universal adoption of these ML models, and benchmarking it. It makes it highly desirable for the community.

Critical weaknesses

While such studies are important and highly valued by the community, I have the reservations on the existing manuscript.

  1. Novelty: such studies are already done by several groups, it is not clear that what do the authors contribute, that has not been discussed previously. Perhaps authors should highlight it.
  2. Clarity: The manuscript in its current form is highly convoluted does not present forth its ideas clearly. It reads like a technical specsheets with numbers after numbers. Perhaps authors should rewrite it, and emphasize the separation of core findings against quantitative enumeration of results.

Minor comments:

  1. Table 1 contains no info other than that MACE uses float64 (why only MACE is in double precision?). Perhaps the table one is not needed?
  2. Right now lot of key figures are referred in the appendix. Perhaps authors might consider them to the main text.
Questions for authors
  1. It says that relaxation of ASE molecules was not carried out as it creates a method dependent structure pathway, which the auhtors want to avoid. Why? I mean reaction pathways will heavily depend on the relaxed structures that each ground truth or the ML models generate. Should it not be relaxed at each point then? This point is unclear.
Reviewer_HnyEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper evaluated MLIPs on 1. local force response 2. static reactive accuracy 3. reaction-path completion.

Critical strengths
  1. paper is clear and easy to follow
Critical weaknesses
  1. there are no new insights. all tests are present in pervious works, and cited.
  2. this paper is just a list of results. there are limited interpretation or discussion for the results, nothing is proposed to improve the failure cases
Questions for authors

None quetions would affact my assessments, this paper needs siginificant changes

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.