- 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
-
Explicit and editable mechanisms with closed-form law fitting.** The outputs are edge-wise force curves 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 . 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).
-
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
-
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).
-
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 (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
-
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.
-
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?