acceptodds
Withdrawn from ICLR 2027

ERPA: Early Prediction of Adsorbate–Metal Distance Changes on Oxide Surfaces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can the initial atomic motion, forces, and energy changes reveal how much the adsorbate–metal distance will change by the end of oxide-surface relaxation? We investigate how these early observations predict the absolute initial-to-final adsorbate–metal distance change during density functional theory (DFT) geometry optimization. Adsorbate identity, surface-metal element, material composition, surface orientation, and initial geometry also help predict this change, making the contribution of the early response the central challenge. We introduce Early Relaxation-Path Endpoint Response Analysis (ERPA) to predict this distance-change magnitude from early displacement, force, and energy descriptors, and assess their predictive value against a static-context reference and a geometry-scrambled control. Descriptors use nominal frames 5 and 10 of OC22 trajectories; shorter trajectories contribute their final stored frame. The scrambled control reassigns geometric increments within adsorbate–anchor groups in each split while retaining the recipient's initial, force, and energy descriptors. In processed held-out surface validation on the geometry-screened subset, the paired representation reduces mean absolute error (MAE) by 0.108173 relative to the static reference and by 0.115808 \AA relative to geometry scrambling. Positive static-control gains extend across O, O, OH, HO, and HO. Force-ratio, energy-drop, and distance-change strata connect these gains to recognizable relaxation responses. A separately refitted strict pre-endpoint evaluation uses observations at indices 4 and 9 only for trajectories with at least 11 frames; on 1022 geometry-screened trajectories in processed held-out surface validation, Path reaches MAE 0.259915 \AA, with gains of 0.094553 and 0.097738 \AA against Static and Shuffle, respectively. Together, these comparisons show that a system's own early geometry, force, and energy response supports prediction of its final adsorbate–metal distance-change magnitude before the stored endpoint.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Oxide-Surface Adsorption, DFT Geometry Optimization, Early Relaxation-Response Prediction, Adsorbate–Metal Distance Change, Controlled Trajectory Comparison

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_LtMJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper claims that early frames of a DFT relaxation trajectory are predictive of the final relaxed structure. They leverage features like early atomic displacements, force ratios, and energy changes from initial frames and train a ridge regression model to predict the scalar change in magnitude of the distance of the closest adsorbate atom to metal oxide in samples from OC22. They show that these additional features improve performance compared to using static features from the initial structure alone.

Critical strengths
  • The paper is on a topic of relevance, namely how to circumvent slow DFT.
  • The authors show significant MAE reduction when incorporating the early relaxation features, suggesting that early relaxation dynamics contain useful information about the final relaxed structure.
  • The authors validate their approach against multiple controls and across different adsorbate species.
Critical weaknesses
  • The method only predicts the absolute magnitude of the change in adsorbate–metal distance. This is of no utility and does not help define the final relaxed structure or energy. It is unclear how this prediction could be used to accelerate DFT calculations or otherwise benefit downstream applications.
  • Although the authors demonstrate improvements over static baselines, it remains unclear how much of this improvement arises from straightforward extrapolation of the observed distance change. The endpoint-matched model performs comparably to or better than the full model, suggesting that additional geometric features contribute little.
  • There is no methodological novelty of this method, which is primarily training a standard ridge regression model on hand-crafted features.
Questions for authors
  • How does predicting the absolute change in adsorbate–metal distance accelerate DFT optimization or benefit downstream applications?
  • How does the method compare against simply using the observed early distance change as the prediction, or extrapolating from early distance changes?
  • What fraction of the final distance change has already occurred within the early frames used by the model? Could the authors report this distribution across trajectories and stratify prediction accuracy accordingly?
Reviewer_Z43COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses the task of predicting the final state of an atomic structure relaxation trajectory from only its first few frames, with a focus on adsorption systems. The core idea is to extract features derived from coordinates, forces, and energies, focusing on the atoms of interest, namely the adsorbate and its neighboring surface metal atom. These features are then used to predict how much the adsorbate-metal distance will change by the final frame. The prediction is made by a simple linear regression model rather than a complex neural network. To validate the effectiveness of the early trajectory information, the authors compare the proposed model against control models that either exclude this information or deliberately corrupt it. In experiments on the OC22 dataset, the authors report that the proposed model achieves lower prediction error than the control models, and that this trend holds for material compositions unseen during training.

Critical strengths
  • The paper addresses a practically motivated problem, as predicting the outcome of a structure relaxation from its early frames could reduce the need for costly full simulations.
Critical weaknesses
  • The experiments lack comparisons with established methods, as most of them compare variants of the proposed approach against each other.
  • The presentation requires substantial improvement.
Questions for authors

I have the following questions for the authors:

  • Most experiments compare variants of the proposed approach against each other. A natural baseline would be a machine learning interatomic potential trained on OC22, such as EquiformerV2. Could the authors provide such a comparison in terms of both accuracy and computational cost?
  • The problem setting appears closely related to that of Schreiner et al. [1], which also aims to predict future states of atomistic trajectories. Could the authors clarify how the proposed work differs from this line of research and provide a comparison?
  • It is unclear why the early response of a trajectory should be sufficient to determine the final distance change. Could the authors provide a theoretical or physical justification for why this task is learnable from early trajectory information?

In addition, I have the following suggestions regarding presentation:

  • The related work in the main text is insufficient, and most of the discussion is deferred to the appendix. The positioning of the work against prior studies should be presented in the main text.
  • Numerical values are listed extensively in the main text, which hinders readability.
  • The method and results are not clearly separated, and the reader is frequently redirected to other sections and the appendix, making the main text difficult to follow on its own.
  • The manuscript is verbose overall, and the writing would benefit from substantial polishing.

Overall, predicting the outcome of structure relaxations from early frames is a practically motivated problem. However, the experiments lack comparisons with established methods, and the presentation requires substantial improvement. I therefore recommend rejection in its current form.


[1] Schreiner et al., "Implicit Transfer Operator Learning: Multiple Time-Resolution Surrogates for Molecular Dynamics." NeurIPS (2023).

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

The authors propose to use inital frame of DFT geometric relaxations to predict the final point. They fit a few ridge regression and MLP and find that it could predict the final point to some degree.

Critical strengths
  1. a lot of careful statistics is evalated and provided.
  2. details is included for reproducibility
Critical weaknesses
  1. The method does not seem to be useful based on the ablations hidden in the appendix. a. from table 8, energy and force features did not help over static features b. from table 6, the trivial baseline is beating the method (signed-distance ridge)
  2. The problem setup is questionable. The lables are optimization artifacts. The model is not learning true physics. OC22 is geometric optimization paths, a. not all of the is converged b. it is not guaranteed the the DFT relax will find the minima c. it is optimizer dependent. The "ground truth" changes with optimizer setting. The mainstream of this field is to train MLIPs to fit the PES, and use that for optimization. At there the model is learning actual physics, not an artifact.
  3. You need to pay for ~10 DFT steps at the start, a full MLIP relaxation could cost less and give a better result.
Questions for authors
  1. how do you explain the trivial baseline was better?

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

Related papers

Discussion (0)

Sign in to comment.