acceptodds
Under review as a conference paper at ICLR 2027

EL-KAN: Learning Sparse Conservative Dynamics from Position-Only Trajectories

Abstract

Learning interpretable particle dynamics from noisy position trajectories requires identifying both which particles interact and the laws governing their interactions. Uneven coverage of interaction distances and unobserved velocities complicate this task, while accurate trajectory fitting alone does not ensure reliable predictions under distribution shifts or interventions. We introduce the Edge-Law Kolmogorov–Arnold Network (EL-KAN), a framework for learning sparse interaction mechanisms in the generated position-only setting, without velocity, force, or graph supervision. EL-KAN represents pairwise interactions through learnable univariate radial potentials whose analytic derivatives produce conservative, equal-and-opposite forces. Weak-form regression transfers temporal differentiation from noisy position observations to known test functions, while sparse graph selection and regularized coefficient refitting identify the interaction structure and governing laws. To address limited radial resolution and discretization bias, EL-KAN combines structured potential candidates, including locally refined radial bases and compact polynomial–exponential dictionaries, with discretization-consistent refitting. Forecasting uses state estimation from an observed trajectory prefix while keeping the learned force field fixed; independent validation selects the potential representation and initialization scheme based on both identification quality and rollout behavior. Experiments on spring, Duffing, and Morse systems under noisy observations and varying sampling resolutions demonstrate effective recovery of sparse interaction structures and nonlinear interaction laws. Compared with the weak-form and anchor controls under the same validation budget, EL-KAN improves long-horizon prediction within and beyond the training distribution and under edge-removal interventions, while retaining explicit, directly editable interaction mechanisms.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Edge-law learning, Interpretable machine learning, Conservative particle dynamics, Weak-form regression, Interaction discovery from positions

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mCa9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies position-only mechanism discovery: from noisy position trajectories alone, recover both a sparse interaction graph and an explicit per-edge force law, assuming conservative pairwise radial interactions with known masses. Two challenges motivate the design: uneven distance coverage makes the right flexibility of the law representation unknowable in advance, and unobserved velocities let initial-state estimation compensate for force-field errors, so prefix fit alone cannot certify mechanism recovery.

EL-KAN assigns each candidate pair a radial potential in a fixed Gaussian basis, differentiated analytically into conservative, equal-and-opposite forces. Coefficients are fitted by weak-form regression (closed-form least squares), and the graph is obtained by ranking edges and refitting top-k masks. Candidate configurations are compared on held-out validation trajectories; a candidate is accepted only if it is no worse than fixed reference models on every criterion and passes a separate confirmation check, otherwise the reference model is used.

On nine synthetic settings, EL-KAN recovers the interaction graph in nearly all runs and substantially improves long-horizon prediction over controls, including under OOD and edge-removal interventions; three public benchmarks serve only as applicability checks.

Critical strengths
  1. Explicit and editable mechanisms with closed-form law fitting.** The outputs are edge-wise force curves ϕij(r)\phi_{ij}(r) and a binary interaction graph. The force-law coefficients are obtained in closed form from a weak-form linear least-squares problem, so law fitting is deterministic and reproducible without gradient-based training. Conservative forces and action–reaction symmetry hold by construction, and counterfactual interventions are defined by setting zij=0z_{ij}=0. Removing a learned edge moves the forecast toward the trajectory of the actual system with that edge removed in all 45 runs (Sec. 4.4).

  2. Evaluation design aligned with the problem setting. The paper argues that low fitting error does not imply accurate mechanism recovery, and evaluates relative law error, OOD generalization, and interventions separately from prediction NMSE. Among the external baselines (the authors’ position-only adaptations), the one with the lowest prediction error (BINN-P) does not output an interaction graph, while those that do (DARI-P, IPSI-P) misclassify more than 4 of 10 candidate pairs on average, near chance level. This shows that trajectory prediction and interaction-structure recovery are distinct, which the evaluation protocol is designed to expose.

Critical weaknesses
  1. The comparisons and datasets are aligned with the model assumptions, with no evidence of generalization beyond them. The external baselines are GNN models that originally require velocity inputs but have been modified by the authors to remove velocity information (-P); these are fixed-protocol adaptations rather than the published methods. Weak SINDy (Messenger & Bortz, 2021) and the kernel method of Lang et al. (2026), which are the most closely related prior methods, are only mentioned in the Related Work section and are not adapted for comparison in the position-only, second-order setting. Furthermore, the simulation settings of the synthetic data (pairwise radial conservative forces, a single force law, and pair-specific weights) exactly match the model assumptions in Equations 1–2. The slope set of the hybrid dictionary also includes both Morse decay rates (1.5 and 3), which the authors themselves acknowledge as a declared functional prior (Sec. 3.1, 3.3), so the Morse recovery results benefit from prior knowledge of the ground-truth functional form. No experiments are conducted under conditions where these assumptions are violated. This can be viewed as a general limitation of linear models that rely on fixed bases and strong structural assumptions (pairwise interactions and distance-only dependence).

  2. The key setting is not validated on real data, and the technical contribution is limited. The central premise of the paper is position-only identification, but all three public benchmark experiments use either force labels (rMD17 and MD22) or velocity and graph labels (NRI) as inputs. Therefore, the weak-form identification (Sec. 3.2) and the state estimation and guarded selection (Sec. 3.4) are never applied to real data, and the authors themselves limit these experiments to demonstrating applicability under their stated supervision. Since there are no results applying the position-only method to molecular trajectories, it remains unclear whether the method works beyond synthetic data. Moreover, the individual components are existing methods: energy differentiation to obtain forces (HNN), a Gaussian radial basis expansion, weak-form identification (Weak SINDy), and the relaxation of zijz_{ij} (Louizos et al., 2018). The novelty lies in combining these components in a position-only setting and in the selection procedures of Sections 3.3–3.4. However, since these procedures are not evaluated on real data, the demonstrated contribution remains limited to synthetic settings.

Questions for authors
  1. The computational components of EL-KAN (fixed basis functions, linear coefficients, weak-form least squares, and sparsity) are closely related to the SINDy family discussed in the third paragraph of Related Work and to the kernel method of Lang et al. (2026). However, the external baselines consist only of position-only adaptations of three GNN models that originally require velocity inputs. Could the authors provide comparisons with the most closely related methods, namely Weak SINDy (with a pairwise library) and Lang et al.’s shared kernel with edge weights (adapted to second-order systems), on the same synthetic datasets? In particular, Lang et al.’s approach has an inductive bias that is better aligned with the synthetic data setting (a shared force law with pair-specific weights) than EL-KAN’s independent per-edge potentials. Without this comparison, it is difficult to determine whether the superiority shown in Table 1 comes from the proposed method itself or from the choice of baselines.

  2. Although rMD17 and MD22 provide force labels, they are derived from molecular dynamics trajectories. Therefore, it should be possible to exclude force labels from the inputs, apply the weak-form identification (Sec. 3.2) and state estimation (Sec. 3.4) using only positions, and report force NMSE against the provided force labels together with rollout errors. This would directly evaluate the paper’s central position-only setting on real data, but such experiments are missing from the submission. Could the authors provide these results, or explain why this evaluation is not feasible?

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

The paper introduces EL-KAN, a framework for learning sparse interaction graphs and conservative interaction laws from noisy, position-only particle trajectories. The method uses learnable radial potentials, weak-form regression, sparse graph selection, and validation-based model selection to recover interpretable dynamics and predict future trajectories.

The authors evaluate EL-KAN on nine synthetic settings involving spring, Duffing, and Morse interactions, as well as three public benchmarks. They report improved long-horizon prediction, graph recovery, and robustness to distribution shifts and edge-removal interventions compared with internal controls and adapted relational-inference baselines.

Critical strengths

None

Critical weaknesses

1. Poor presentation and unclear methodological contribution

The manuscript is extremely difficult to follow and requires substantial rewriting. The Abstract and Introduction fail to clearly motivate the problem, explain its challenges, or establish why existing approaches are insufficient. Instead, the Introduction almost immediately lists existing methods without adequately introducing the underlying problem. Consequently, the contributions in lines 73–83 remain difficult to understand, particularly for readers unfamiliar with this research area.

These problems persist throughout Section 3. The methodology introduces numerous components without sufficiently explaining their motivation or distinguishing established techniques from novel contributions. Moreover, Sections 3.1–3.4 contain very few citations, with Sections 3.2 and 3.4 having none. This is particularly concerning given the reliance on established ideas such as conservative potentials, weak-form identification, and sparse regression.

The manuscript also contains numerous editorial errors, including unresolved references ("equation equation 1"), repeated words ("Let Let", line 303), and awkward formulations. Although individually minor, their frequency further undermines the readability of the paper.

Overall, the presentation is not merely a stylistic issue: it substantially obstructs the evaluation of the scientific contribution and makes it difficult to determine what is genuinely novel about EL-KAN.

2. Insufficient experimental validation and reporting

The experimental evaluation does not convincingly support the paper's main claims.

First, although EL-KAN is presented as a position-only identification framework, the public benchmarks use different supervision assumptions: rMD17/Aspirin and MD22/DHA are force-regression tasks, while NRI/Springs-5 uses velocities and graph labels (Section 4.1). Moreover, no external baselines are evaluated on these public datasets. The external comparisons are restricted to synthetic five-particle systems using adapted versions of existing methods, with potentially different training protocols.

Second, Table 1 reports results over five seeds (three for external baselines) without standard deviations or confidence intervals, making it difficult to assess the reliability of the substantial reported improvements.

Finally, Appendix C (lines 733–736) indicates that the radial basis range for public benchmarks is computed using training, validation, and test snapshots, raising potential concerns about test-set information entering preprocessing.

These limitations, together with the restricted scale of the synthetic experiments, weaken the evidence for the method's generality and superiority over existing approaches.

Overall assessment

While the paper reports promising results, its poor scientific exposition prevents a clear assessment of the methodological contribution, and the experimental evaluation leaves important questions unresolved. Addressing these concerns would require a major revision, including substantial rewriting and stronger empirical validation. I therefore recommend rejection in its current form.

Questions for authors

See critical weaknesses.

Reviewer_tHY5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a scheme to learn the interaction relationships of a set of point clouds. It uses kernel function estimation, then selects valuable interaction edges, and finally performs multi-candidate verification to estimate the interaction forces between particles.

Critical strengths
  1. The author's overall pipeline is quite concise and clean.
Critical weaknesses
  1. Weak-form regression, sparse graph selection, radial basis functions, shooting, and validation set model selection are all existing techniques and lack innovation.

  2. The experiments contain some leakage. Appendix C indicates that the authors calculated the radial basis range of the public molecular data using the positions of train+validation+test; however, the radial basis range of the validation test is actually unknown.

  3. The authors also need to explain why they provided 11 frames of history for EL-KAN while the baseline only provided 5 frames, which is highly unfair.

Questions for authors
  1. Could you elaborate on the case of basis function radius minus OOD?

  2. If the number of particles increases to 1000 or more, is this method still applicable? I understand the complexity might be N^2.

  3. Regarding the baseline comparison, an explanation is needed as to why 11 frames of historical data aren't consistently provided.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.