acceptodds
Withdrawn from ICLR 2027

Finite-Propagation-Aware Multiresolution Neural Operators for Structural Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning structural-dynamics surrogate models requires reasoning over irregular meshes and sparse sensor measurements while respecting the finite wave-propagation time for loads to arrive at different spatial locations. Standard neural operators can capture long-range spatial interactions, yet unconstrained message passing can produce unphysical early responses that violate wave-causality constraints. We introduce FPMR-NO, a finite-propagation-aware multiresolution neural operator that combines causal load-history encoding, wave travel-time gates, and coarse-to-fine graph aggregation. The model propagates separate latent messages corresponding to primary/compressional waves (P-waves), secondary/shear waves (S-waves), and mixed wave modes across graph edges only once the source-to-edge wave arrival condition is satisfied. This multiresolution pooling shortens computational paths without introducing unphysical zero-time shortcuts. On the DeformingPlate stress task, FPMR-NO achieves the lowest mean normalized relative L2 error among the evaluated methods, with a 2.4% relative improvement over AMG and a 1.7% improvement over Transolver. On real ETH four-story impact tests, it reduces macro-channel normalized root mean squared error (NRMSE) to 0.86, a 14.1% error reduction relative to the strongest baseline. Controlled graph-wave experiments further show a 71.8% reduction in relative L2 error and a 13.3-fold reduction in pre-arrival energy against an ungated counterpart.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, structural dynamics, finite propagation, wave propagation, graph neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_uyuEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The submission proposes FPMR-NO, a finite-propagation-aware multiresolution neural operator designed specifically for structural dynamics. Standard GNO / GNN often allows immediate message passing across established edges. This design creates unphysical shortcuts because computational connectivity outpaces actual physical wave reachability. The authors solve this by embedding finite wave propagation as a strict architectural prior.

Critical strengths
  1. The reviewer appreciated that the formalization is easily verifiable. Assumption are clear and the propositions are generally mathematically correct.

  2. The gate design is interpretable and permits continuous speed adjustments. The model is highly lightweight. It uses 0.21M parameters compared to AMG's 2.38M.

  3. The architecture successfully recomputes material travel times at each graph scale. It applies differentiable arrival gates to separate P-wave, S-wave, and mixed message branches.

Critical weaknesses
  1. One of biggest weaknesses of this method is that the gate enforces first-arrival support, not finite-propagation causality. Also, the performance gain is relatively marginal. For example, the accuracy improvement over the Transolver baseline on the Plate task is relatively marginal at 1.7%.

  2. DeformingPlate is quasi-static. It lacks explicit time steps, material constants, and true wave speeds. The propagation gate merely acts as an unspecified structural graph bias here.

  3. The ETH task predicts sparse accelerometer responses; only ablates the transfer initialization and geometry residual. A true elastodynamic multiscale benchmark with gating ablations is completely missing.

  4. The ETH dataset provides sparse accelerometer channels instead of full-field stress or displacement measurements.

  5. Long-horizon temporal stability remains a challenge. Errors clearly accumulate during autoregressive rollout on the DeformingPlate task.

Questions for authors

Which softness parameter η\eta do the main tables use?

Does the ungated baseline still receive arrival times?

What are the load inputs and wave speeds on Plate?

If this is the official AMG setting, AMG originally reported an error lower than your 0.0260. If you align with their numbers, your margin over AMG reverses.

How is NRMSE normalized on ETH?

Why does Figure 16 drop to roughly 0.50 at horizon 4096 when Table 2 explicitly reports 0.86?

If you run a two-pulse test, does the predicted response change when only later source loads are perturbed?

Reviewer_7vfjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FPMR-NO, which incorporates wave travel-time gating, causal load-history encoding, and multiresolution graph message passing for structural dynamics prediction. The framework is evaluated on DeformingPlate, experimental structural impact responses, and controlled wave-propagation benchmarks.

Critical strengths
  • The distinction between computational graph connectivity and physical wave propagation is well motivated and addresses a meaningful limitation of unconstrained graph-based operators.

  • The controlled wave-propagation experiments provide direct evidence that travel-time gating can suppress unphysical pre-arrival responses.

Critical weaknesses
  • The DeformingPlate task lacks the material parameters and temporal information required to identify physical wave arrival times, so the propagation gate acts primarily as a structural prior. In the ETH experiments, the individual contribution of propagation gating is not sufficiently isolated from the transfer initialization and modal response modeling. Therefore, the reported improvements cannot be attributed conclusively to finite-propagation awareness. The contribution of finite-propagation constraints is not convincingly isolated in the real-world evaluations.

  • Soft arrival gates permit nonzero pre-arrival message transmission, while multiresolution gating does not necessarily enforce exact arrival-time constraints for individual fine-scale nodes. The relationship between the stated causality guarantees and the implemented prediction operator requires further clarification. The claimed propagation constraints are stronger than the guarantees provided by the implementation.

Questions for authors
  • On the ETH dataset, what is the isolated effect of removing propagation gating while retaining the transfer initialization, modal response head, and other architectural components?

  • Can the method be evaluated on a transient finite-element benchmark with known material properties and wave arrival times? Direct measurements of arrival-time errors and sensitivity to wave-speed misspecification would help establish the physical validity of the proposed constraints.

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

The paper studies structural-response prediction and proposes that information exchange in a graph model should account for the time needed for a physical disturbance to travel. The proposed method, FPMR-NO, combines past-load encoding, arrival-time gates, and graph processing at multiple resolutions. Experiments cover DeformingPlate stress prediction, ETH laboratory impact responses, and controlled synthetic tasks. Reported results include lower average errors on the two main benchmarks and reduced responses before wave arrival in the synthetic setting. The paper also analyzes which restrictions on information flow are enforced by the architecture.

Critical strengths
  • I think the motivation is reasonable: a model should account for the time needed for a disturbance to reach another location.
  • Using arrival times to control message passing is a simple and understandable design.
  • Results on the controlled wave task show lower prediction error and less response before arrival than the ungated model.
  • Code is publicly available.
Critical weaknesses
  • I am not convinced that the two main benchmarks demonstrate the benefit of the proposed propagation constraint. In the released implementation, the Plate gates are effectively always open, while the ETH model does not use the proposed graph gates. Although Section 3.5 describes task-specific adaptations, these results do not establish whether enforcing travel time improves prediction.

  • I find the Plate gating ablation difficult to interpret. Table 4 reports a 3.01% increase in error after removing the gate, even though the released configuration already leaves it open. Without an explanation of this discrepancy, I cannot take that result as evidence that gating helps.

  • On ETH, the baseline results leave me uncertain about how strong the comparison is. Their main errors are close to or worse than the reference level obtained by predicting each test channel’s mean, and the paper does not show how well its own transfer initialization performs before further training. Reporting that simpler comparison would help clarify what the learned model adds.

  • I think the novelty needs to be explained more clearly in relation to earlier work. Multiscale graph operators such as AMG and temporal gating methods such as Causal-DeepONet already cover parts of the design, although neither is identical to the proposed method. Combining these ideas with spatial arrival times could be useful, but the current experiments leave me unconvinced of the added value of this combination.

Questions for authors

Please check weakness

Reviewer_7VJZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors present a neural operator for predicting how structures respond to applied forces, such as stress in a deforming plate or vibrations caused by an impact. Their main idea is to align message passing inside the neural network with physical wave propagation, so that information passes between nodes only once a physical wave could have reached them. To achieve this, they construct multiresolution graphs, estimate wave travel times along graph edges, and use gates to control message passing based on these times. They also incorporate the history of applied forces. The method is evaluated on DeformingPlate, real-world impact measurements from an ETH structural frame, and synthetic wave-propagation experiments, where it generally outperforms the tested baselines.

Critical strengths
  • The physical motivation is sound and relatively clear.
  • On controlled synthetic data, the gating mechanism clearly helps.
Critical weaknesses
  • According to the ablations, the mechanism does not seem to help much on real problems. Moreover, the improvement over the baselines is quite small (under 3% for Table 1) or lies within a regime where all models perform extremely poorly (Table 2; see also Q1 below).
  • The released code shows that FPMR-NO receives hyperparameter tuning, while comparable tuning is not documented for the baselines. Furthermore, the larger improvements on ETH may be driven by the specialized transfer initialization rather than by the proposed propagation gating.
  • The estimated arrival times depend on the constructed graph and its coarsening, and the authors do not show that they correspond to actual physical wave arrival times. It is difficult for the reader to judge whether the proposed method truly incorporates physical travel-time constraints or merely introduces a graph-dependent inductive bias.
Questions for authors
  1. In Table 2, the NRMSE is generally around 100 for most models, and according to the authors, these numbers are to be read as mantissas with a scale factor of 10^-2. This means the NRMSE is of order 1, which would make all models extremely poor predictors, with the authors’ model being better, but only marginally so. Is this correct? I find it strange that established models would perform so poorly on these problems.
  2. In Table 2, Poseidon is listed with 0.11M parameters, but to my knowledge no model of this size exists; the smallest Poseidon model (Poseidon-S) has 21M parameters. Where does this model come from?
  3. In general, how were the baselines configured and tuned (especially given the performance in Table 2)?

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

Related papers

Discussion (0)

Sign in to comment.