acceptodds
Under review as a conference paper at ICLR 2027

Learning Mode-Resolved Response Computations across Systems and Forcing

Abstract

Predicting complete responses to external forcing supports the design and assessment of physical systems under changing loads. Generalizing across systems requires adapting to both new excitation and different dynamic scales. We present a system-conditioned framework for full-event prediction. Its Mode-Resolved Response Operator (MRRO) uses supplied modal frequencies to condition kernel coefficients and place finite bases at each mode's scale. Propagated latent features return to node coordinates for nonlinear coupling on the query's physical-time grid, with the output anchored to a computed reference. Weights are shared within each domain and trained separately across domains. Matched experiments control physical information, phase and capacity, while assessing dependence on optimization budget. On new systems and forcing, MRRO reduces acceleration error by 45.28% and 27.83% relative to F-FNO adaptations in hysteretic and torsional networks. A joint building model covers earthquakes and six applied-force programs; five paired training runs yield 14.15% lower error than common-scale kernels on 582 responses from 128 held-out families. A proposal-conditioned reaction–diffusion model and calibrated folding-wing experiments examine extensions of the kernel construction. Frozen mechanical predictors also compare candidate actions without goal-specific training. These findings show that physical-scale allocation matters for transferable response learning.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural operators, Scientific machine learning, Cross-system generalization, Forced dynamics, Modal scale allocation

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

This paper proposes a neural network model that predicts responses given external force. The proposed method, mode-resolved response operator (MRRO), uses modal frequencies for kernels. The neural network transforms the hidden states considering modal information. In the experiments on five types of systems, such as mechanical, torsional, and joint buildings, the proposed method achieves better prediction performance than the existing operator learning methods. By the ablation study, the effectiveness of the kernels for each mode is demonstrated. The authors also evaluate the proposed method on the performance of selecting action.

Critical strengths

The neural network design considering mode information is interesting.

The experimental results show the effectiveness of the proposed method in different systems. The ablation study demonstrates the effectiveness of the component in the proposed method.

Critical weaknesses

The motivation of the design of the proposed method is not clearly described. The authors describe equations without the explanation why they are effective.

The novelty of the proposed method is unclear. The modes of the system have been incorporated in the neural networks of the existing methods. More detailed discussion and comparison with the existing related methods are needed.

The proposed method is based on kernel basis. It is unclear the advantage of this approach over existing methods (e.g., methods listed in “Relation to existing operators”).

The presentation can be improved. Figure 1 is not referred to in the main text. It would be better to define all abbreviations at their first occurrence in the main text, even if they have already been defined in the abstract. There is not sufficient explanation on E001 in Section 2.

Questions for authors

Please see the weaknesses。

Reviewer_Ff8oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MRRO for predicting complete responses of physical systems under changing forcing. Temporal-kernel bases are placed at the scale of each supplied modal frequency, and the modal outputs are combined in node coordinates to model nonlinear interactions. The authors test the approach on mechanical and torsional networks, buildings, reaction-diffusion systems, and measured wing responses, with separate models trained for each domain. The main question is whether mode-specific basis placement helps beyond simply supplying modal information to the network.

Critical strengths
  1. The common-scale comparison is the most convincing part of the paper. Both models receive the same physical information and use the same number of basis functions, so the comparison tests basis placement rather than access to extra information. The phase and capacity controls in Section 5 further support this interpretation.

  2. The mechanical and torsional tests include both new systems and new forcing. The building results also show a consistent advantage over common-scale kernels across five training runs (Table 3). These results provide useful evidence that the idea is not limited to fitting responses from a single system.

Critical weaknesses

My main reservation is inference cost. MRRO uses fewer parameters than enhanced C-PhysFNO, but its cached queries are slower: 21.77 versus 12.40 ms for applied loads, and 23.57 versus 22.45 ms for earthquakes (Section 4.2). The accuracy gains are substantial, so this is not a reason to reject the method on its own. Still, the current results establish an accuracy benefit, not an advantage at a matched inference-time budget.

Questions for authors

Would a lower-cost MRRO configuration retain the accuracy advantage at a query time comparable to enhanced C-PhysFNO?

A runtime-matched comparison would help assess the method for applications that require many repeated predictions.

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

This paper proposes the Mode-Resolved Response Operator (MRRO), a system-conditioned neural operator for predicting the full temporal responses of physical systems under external forcing. The key idea is to incorporate modal frequencies into the construction of temporal kernel bases, allowing different dynamical modes to be represented at their corresponding physical scales. The resulting latent features are mapped back to physical coordinates to capture nonlinear interactions between modes.

The authors evaluate MRRO on several physical systems, including hysteretic and torsional networks, building dynamics, reaction–diffusion systems, and folding-wing experiments. The results demonstrate improved predictive accuracy over existing neural operator baselines, particularly in generalization to unseen systems and forcing conditions. The paper argues that allocating kernel bases according to physical modal scales can improve transferable response prediction.

Critical strengths

The proposed method achieves strong empirical performance, outperforming several existing operator learning methods on the evaluated benchmarks.

Critical weaknesses

Methodological justification. My main concern is the insufficient connection between the proposed methodological motivation and the actual limitations of existing operator learning methods.

The paper argues that allocating finite temporal bases according to mode-specific physical frequencies can improve response prediction. I acknowledge that Section 3.2 provides a reasonable motivation for this design, particularly through the local coordinate-density argument in Eq. (5).

However, it remains unclear to me whether the identified limitation of common-scale basis allocation is indeed a fundamental limitation of existing operator learning approaches. The related work discusses several neural operator architectures, but does not sufficiently explain how their modeling assumptions lead to the specific problem that MRRO aims to address.

Furthermore, while Eq. (5) motivates mode-specific basis placement, its implications for approximation accuracy in nonlinear, coupled dynamical systems remain unclear. I would appreciate a more explicit discussion connecting this mathematical observation to the reported empirical improvements.

I believe the paper would benefit from a more critical examination of existing approaches and a clearer explanation of the scientific insights provided by the proposed architecture, beyond its empirical performance.

Presentation and clarity. I find the exposition unnecessarily dense, particularly in the related work and methodology sections. The paper introduces numerous architectural components and mathematical definitions in rapid succession, often without sufficient intuition or explanation of their individual roles.

The related work section is especially difficult to follow as a scientific narrative. Although several existing operator learning methods are cited, their underlying assumptions, limitations, and relationships to the proposed method are not adequately discussed. As a result, readers are left to infer much of the motivation for the proposed architectural choices.

I also find Figures 1 and 2 overly crowded, making it difficult to identify the essential methodological contributions.

While the central idea of mode-resolved basis allocation is reasonably clear, I believe the manuscript would benefit substantially from a more structured exposition that prioritizes conceptual understanding over implementation details.

Questions for authors

Q1. What specific structural limitations of existing operator learning methods motivate the proposed mode-resolved basis allocation? How do these limitations arise from their underlying modeling assumptions?

Q2. How does the local coordinate-density argument in Eq. (5) translate into improved approximation accuracy, particularly for nonlinear systems with strongly coupled modes?

Q3. Is the observed improvement specifically attributable to mode-resolved physical scaling, or could similar improvements be achieved using alternative adaptive or multiscale basis constructions under the same information and computational budget?

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

This paper introduces MRRO, a neural model that predicts the full time response of structural and physical systems to external forcing, such as earthquakes and wind loads. MRRO computes the modal frequencies of each system from its mass and stiffness matrices and uses them to place a fixed set of temporal kernel functions at the scale of each mode, instead of one scale shared by all modes. On five datasets, MRRO has lower error than adapted neural operator baselines, and on buildings it reduces the error by 14.15% compared with a version that uses one common scale.

Critical strengths

The main design choice is tested with controls that keep the input information and the number of parameters the same (Section 5), and the building comparison uses five training runs (Table 3). The study covers several physical domains, including measured wind-tunnel data of folding wings, and reports the cases where the method does not help, such as the reaction-diffusion task.

Critical weaknesses

The evidence for the main idea is unclear. Per-mode placement clearly helps on torsion and buildings, but the mechanical gain is about 5% with two runs whose results overlap (Figure 5a), the two versions are tied on reaction-diffusion, and Figure 5 reports validation rather than test error.

Several evaluation issues weaken the results. The test sets for buildings, fields and wings were seen during development (Table 1), earthquake records are not fully separated between splits, and more than half of the planned building simulations failed and were removed (Section 2).

The comparison with external baselines is not reliable. Many baselines fail to fit the training data and give identical errors (e.g., 3.1399 for GNOT and UPT, 0.5728 for LNO, GNOT and UPT, Table 2), and most external results use a single run, so the large gains over F-FNO depend on the quality of these adaptations.

The method is not one framework across domains. The reaction-diffusion task uses a different network in which the per-mode part adds only 0.7% of the parameters, and the wing predictions show little skill (relative error 0.93–0.95, only 3.93% better than a simple reference built from the initial record).

The presentation needs substantial improvement. The text is very dense, and many terms and design choices are only defined in the appendix, so the main text is difficult to follow on its own.

Questions for authors
  1. What are the test errors of the common-scale and per-mode versions on the mechanical and torsional datasets, over more than two runs?

  2. How do the building results change on a test set that was not used during development, with earthquake records fully separated between splits?

  3. Why do several baselines give identical errors, and were these adaptations checked on a simple case before the comparison?

  4. Why is the torsional validation error in Figure 5 (about 0.05) much lower than the test error in Table 2 (0.2103)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.