acceptodds
Under review as a conference paper at ICLR 2027

SONAR: Sensor-Conditioned Neural Operators with Adaptive Spectral Rotation for Full-Field Prediction on Meshes

Abstract

Existing physical-field prediction paradigms include autoregressive simulators and in-situ sensors reconstruction, which rely on complete prior field data or dense in-domain sensors and thus cannot be deployed under practical equipment operating conditions. Real engineering systems only provide sparse low-dimensional time-series observations. These observations mismatch target physical fields in both units and spatial locations, making it hard to infer high-dimensional field evolution from limited sensor measurements. This raises a critical challenge for high-precision physical field prediction of edge-side equipment: revealing how sparse heterogeneous observations propagate in high-dimensional field spaces and regulate global field distributions. Therefore, we define a more rigorous edge-side physics-AI task: causal next-step full-field prediction on unstructured meshes from short windows of heterogeneous observed state sequences. Duhamel–Volterra and observability analysis show that the predictable component of the field is fully spanned by causal convolutions of the observed states, unobservable field energy sets an error floor, and the optimal predictor admits a conditional spectral decomposition. We instantiate this framework in SONAR: observation history drives the rotation of mesh spectral basis fields generated by deep message passing and architecturally Gram-whitened; sensor-derived coefficients synthesize the global field, and a curvature-guided active-set residual repairs localized high-frequency errors. Experiments on elastic-beam, heat-plate, and mechanical-leg benchmarks with ablation, uncertainty, and robustness analyses show that SONAR achieves the best accuracy among compared methods under a matched information budget.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operator, physics-informed machine learning

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

The paper proposes a method for predicting future states of a physical system from sparse heterogenous observations.Various techinces are used such as Gram-Whitening, message passing, and active learning; however, it is very difficult to understand how these come together to make the proposed method, primarily because the writing is overly complicated and made unnecessarily technical and jargon-laden. This makes the main contribution highly obscure.

Critical strengths

The overall problem the paper is trying to assess, prediction of future states of physical systems from sparse heterogeneous observations, is a good problem to try and tackle.

Critical weaknesses

The paper is very poorly written with very confusing and non-standard terminology. This makes it very difficult to understand the high-level approach of the paper. On this basis alone I believe this paper should be rejected.

Questions for authors

None. I strongly advise a complete re-write and keeping the reader in mind and emphasizing clear language.

2 comments

Thank you for the direct and candid assessment. We sincerely apologize that the initial manuscript was overly dense, used unnecessarily technical terminology, and made it difficult to understand how the components formed a coherent method. This prevented the core contribution and the empirical evidence from being clear. Your criticism is well taken.

We have undertaken a complete rewrite with the reader as the primary consideration. We respectfully ask you to evaluate the revised manuscript as a substantially rewritten presentation of the work rather than as a collection of local textual edits.

Response to the request for a complete rewrite and clearer language

1. The problem and contribution are now stated before technical details

The revised introduction first explains the input contract: the model receives a short history of heterogeneous system-level observations and a mesh, but no current or past target field. It then contrasts three representations in plain terms:

  1. Fixed-basis methods change only coefficients within a shared spatial subspace.
  2. Direct mesh decoders change node features without an explicit low-rank interface.
  3. SONAR generates the spatial basis itself for each observation window and then predicts coefficients within that basis.

This leads directly to the central contribution: state-conditioned basis generation that separates per-sample spatial subspace construction from coefficient prediction. The abstract, introduction, related work, and conclusion now use the same consistent description.

2. The method section now follows the computation in a reader-oriented order

Section 3 is reorganized into a task definition, the core idea, basis generation, coefficient prediction, localized residual correction, training, and a short theoretical interpretation. The central computation is introduced in one equation before the architectural details:

\hat{Y}_{\mathrm{glob}}=\Phi_{\theta}(G;z_s)a_{\psi}(z_s,z_G),\qquad \hat{Y}=\hat{Y}_{\mathrm{glob}}+\hat{Y}_{\mathrm{res}}.$$ The respective roles are now explicit. The recurrent encoder summarizes observation history. The mesh network generates basis fields for the current state. The coefficient head determines how strongly each basis field is expressed. Whitening normalizes the basis by construction. The residual branch corrects localized structure outside the global low-rank approximation. ### 3. Non-standard and potentially confusing terminology is explicitly defined The revised manuscript states that adaptive spectral rotation means a change in learned basis directions across observation states. It does not refer to eigenvectors precomputed from a fixed mesh operator. It also states that the active node set is a subset of mesh nodes used for localized computation and is unrelated to active learning. These clarifications directly address the terminology that made the initial version difficult to interpret. ### 4. Complex derivations no longer interrupt the main narrative The detailed assumptions, propositions, and proofs are consolidated in Appendix A. The main paper retains only a concise Section 3.7 explaining how the analysis motivates recursive observation encoding, state-dependent bases, and whitening. It also states explicitly that the theory does not establish global optimality or attainment of the observability limit on the benchmarks.

5. The experiments are organized around clear research questions

The evaluation now proceeds through five questions: accuracy under the sensor-only contract, mechanisms behind the improvement, computational cost, mesh generalization, and uncertainty and observation robustness. Baseline adaptations are described in a dedicated Appendix D.3, and complete five-seed results are provided in Appendix E.

The revised evidence includes several additions that make the contribution easier to assess:

  • A learned shared-basis control isolates the effect of sensor-dependent basis generation.
  • Cross-resolution and unseen-geometry tests evaluate generalization directly.
  • Parameter counts, training cost, memory, inference speed, and whitening cost are reported.
  • Go1Leg-Real provides 4,500 independent conditions with physical robot sensor inputs.

For example, the key basis control in Table 4 is now straightforward to interpret:

Basis representationBasis varies by sampleElasticBeam-Syn NMSEGo1Leg-Real NMSE
Fixed PODNo0.0612 ± 0.00340.3027 ± 0.0126
Learned shared graph basisNo0.0435 ± 0.00380.2989 ± 0.0082
SONAR conditioned basisYes0.0249 ± 0.00090.1673 ± 0.0053

The control retains the same generator, coefficient head, rank, whitening, residual branch, and training protocol, but removes observation input to the basis generator. The result therefore tests the central mechanism directly rather than relying on architectural description alone.

On Go1Leg-Real, SONAR achieves NMSE 0.1673, RelL2 0.3129, and VM-MaxRE 9.8, compared with the strongest baseline results of 0.2019, 0.3458, and 14.2, respectively. The relative reductions are 17.1%, 9.5%, and 31.0%. The revised paper also reports where effects are small or statistically inconclusive instead of making uniformly strong claims.

We again apologize for the initial presentation and thank the reviewer for prompting a fundamental improvement of the manuscript. We hope that the revised paper now makes the problem, method, contribution, and evidence accessible and coherent. We would be grateful if you could read the revised version and reconsider its presentation, informativeness, and soundness, including your overall evaluation, in light of these extensive changes.

Reviewer_fXwsOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies next-step full-field prediction on an unstructured mesh from a short history of low-dimensional sensor observations, in the setup where previous target fields are not available. It proposes a method which encodes the sensor sequence with a GRU, generates state-conditioned mesh basis fields using deep message passing, applies whitening, predicts basis coefficients, and adds a curvature-selected residual correction. Experiments are conducted on three small datasets, and report the best average performance compared to baselines. Those baselines were modified to fit the experimental settings of the studied problem.

Critical strengths
  • The paper targets an important and practical problem, that of predicting mesh-based physical fields when previous full-field states are not available.
  • The different parts of the method (combines temporal conditioning, mesh message passing, low-rank synthesis, and localized residual correction) are combined in a reasonable manner. Ablation shows that the final method gains from its constituents.
Critical weaknesses
  • Experiments are conducted on very small datasets.
  • It seems that claims regarding good performance with real sensor measurements are not evaluated.
  • No computational cost is reported.
  • The competing methods are were modified to satisfy the experimental setting. The comparisons measure performance of specific adaptations, and not the published methods in their original settings
  • Performance gains against MeshGraphNets and LSTM-GNN are not substantial, even though these methods seem to have been changed substantially to fit the experimental settings.
Questions for authors
  1. Can you report parameter counts, training cost, memory use, and inference speed in the comparisons?
  2. How does the method perform when using real sensor measurements?
  3. The selection of the modifications of the prior methods for being adapted to the experimental setting needs to be better described.
  4. How does the method perform when using larger datasets.
3 comments

We sincerely thank the reviewer for recognizing the practical importance of the task and for providing precise criticisms regarding dataset scale, real sensor evaluation, computational cost, baseline adaptation, and effect sizes. We have addressed each point in the revised manuscript. The main changes are the new 4,500-condition Go1Leg-Real benchmark with physical robot sensor inputs, complete computational profiling, a detailed Appendix D.3 describing every baseline adaptation, and five-seed statistical comparisons that distinguish substantial improvements from small or inconclusive ones.

Critical Weakness 1: Very small datasets

We agree that the original 500-condition mechanical-leg dataset limited the strength of the empirical evaluation. The revised paper replaces it with Go1Leg-Real, which contains 4,500 independent operating conditions, a ninefold increase in the number of independent conditions. Each condition contains 300 time steps of physical robot sensor measurements, and the model uses a 50-step window to predict the next-step field.

The dataset is divided into 3,375 training, 562 validation, and 563 test conditions. Splits are made over independent operating conditions rather than over windows drawn from shared trajectories. The full results under this larger dataset are provided in Q2 and Q4 below and in Tables 2 and 11 of the revised manuscript.

Critical Weakness 2: Lack of evaluation with real sensor measurements

We agree that the original evaluation did not adequately support this aspect of the problem motivation. The new Go1Leg-Real benchmark uses 17 channels of physical robot sensor histories, including joint positions, joint velocities, joint torques, foot reaction forces, hip-mounted inertial accelerations, and trunk gyroscope rates. No internal full-field observation is provided to the model.

For clarity, the dense stress labels are not claimed to be fully measured ground truth. Partial nodal stress measurements are combined with a simulation-based digital twin, and a reconstruction algorithm recovers the six-component stress tensor at all 22,515 nodes. This gives a hybrid real-sensor/simulation-field evaluation of the complete input-to-field pipeline. Section 4.2 and Appendix B.3 now state this distinction explicitly.

Under this evaluation, SONAR achieves NMSE 0.1673, RelL2 0.3129, and VM-MaxRE 9.8, compared with 0.2019, 0.3458, and 34.7 for LSTM-GNN-adapted, and 0.2927, 0.5073, and 39.2 for MeshGraphNets-adapted. SHRED obtains the second-best VM-MaxRE of 14.2. The complete comparison is included under Q2.

Critical Weakness 3: No computational cost reported

We added both end-to-end and component-level cost measurements. Table 6 reports parameters, training GPU hours, peak training memory, peak inference memory, and batch-one latency under a unified protocol. Table 10 additionally isolates whitening forward time, backward time, memory, and full-model latency across basis ranks.

For example, SONAR has 2.4M parameters, requires 6.8 training GPU hours, uses 18.6 GB peak training memory and 3.2 GB peak inference memory, and has 145 ms batch-one latency on ElasticBeam-Syn. The complete comparative table is provided under Q1. These additions make the accuracy and efficiency trade-off explicit.

Critical Weakness 4: Baselines measure specific adaptations rather than the original methods

We agree with this interpretation. The revised manuscript now states explicitly in Sections 2, 4.2, and 5 that all comparisons concern task-adapted implementations under a common information contract, not the published methods in their native input settings.

This adaptation is necessary because the original input contracts of several methods require the previous full field, while our task strictly prohibits current or historical target-field observations. We therefore remove field-history inputs and supply the same observation information through method-appropriate interfaces. Appendix D.3 and Table 9 now specify every modification, the preserved architectural components, and the implementation basis. Appendix D.2 also reports the validation-based hyperparameter search for every baseline.

We have revised the claims so that they do not imply superiority over the original methods in their native settings.

Critical Weakness 5: Gains over MeshGraphNets and LSTM-GNN are not substantial

We agree that some effects in the original synthetic evaluation were small, and the revised paper now acknowledges them explicitly. ElasticBeam-Syn RelL2 differs by only 0.4% with p=0.872p=0.872, and HeatPlate-Syn VM-MaxRE differs by 1.1% with p=0.054p=0.054. We do not present these as statistically confirmed advantages.

The larger Go1Leg-Real evaluation provides a clearer separation:

MetricSONARMeshGraphNets-adaptedLSTM-GNN-adaptedSHREDReduction against strongest baseline
NMSE0.16730.29270.20190.283117.1%
RelL20.31290.50730.34580.49539.5%
VM-MaxRE9.839.234.714.231.0%

The corresponding pp-values are <0.001, 0.002, and <0.001. Thus, the revised results support a meaningful advantage on the complex real-robot-informed benchmark while retaining cautious conclusions for the smaller synthetic effects.

Q1: Can you report parameter counts, training cost, memory use, and inference speed?

Yes. The revised Table 6 reports all four quantities under the same hardware, numerical precision, batch size, and synchronization protocol on ElasticBeam-Syn. Inference latency is the median of 100 synchronized forward passes after warm-up.

MethodParametersTraining GPU hoursPeak training memoryPeak inference memoryBatch-one latency
SONAR2.4M6.8 h18.6 GB3.2 GB145 ms
MeshGraphNets-adapted2.1M4.2 h14.8 GB2.4 GB95 ms
LSTM-GNN-adapted2.0M4.8 h15.4 GB2.6 GB110 ms
SHRED0.7M0.35 h1.2 GB0.45 GB8.5 ms

SONAR is more computationally expensive than the compared baselines, while SHRED has the lowest cost but higher field-prediction error. We now state this trade-off directly. Appendix D.6 further reports the cost of Gram whitening on Go1Leg-Real: at the default rank R=64R=64, whitening takes 0.85 ms forward and 1.90 ms backward, with 182 MB peak memory.

Q2: How does the method perform when using real sensor measurements?

We added Go1Leg-Real to evaluate this question directly. The model inputs are physical robot sensor histories, while the dense stress targets are reconstructed from partial nodal measurements and a simulation-based digital twin. This is therefore an evaluation with real sensor inputs, but not a claim of fully measured dense field ground truth.

The complete Go1Leg-Real results are shown below. Entries are mean ± one standard deviation over five training seeds, as reported in revised Tables 2 and 11.

MethodNMSE ↓RelL2 ↓VM-MaxRE ↓
DeepONet0.8172 ± 0.00710.8021 ± 0.0096298 ± 16.2
MeshGraphNets-adapted0.2927 ± 0.00830.5073 ± 0.004339.2 ± 5.8
BSMS-GNN-adapted0.3420 ± 0.01270.4927 ± 0.016239.7 ± 4.6
GNOT0.5691 ± 0.01020.5194 ± 0.011398.1 ± 20.1
SHRED0.2831 ± 0.00310.4953 ± 0.002514.2 ± 0.57
S-DeepONet0.8933 ± 0.00750.7825 ± 0.0081319 ± 8.4
Transolver0.4129 ± 0.01320.4360 ± 0.010975.4 ± 12.0
Erwin0.3586 ± 0.00980.5097 ± 0.0097154 ± 18.3
LSTM-GNN-adapted0.2019 ± 0.00620.3458 ± 0.012734.7 ± 2.9
SONAR0.1673 ± 0.00530.3129 ± 0.00959.8 ± 0.49

SONAR achieves the best result on all three metrics. Relative to the strongest baseline for each metric, the reductions are 17.1% for NMSE, 9.5% for RelL2, and 31.0% for VM-MaxRE, with pp-values of <0.001, 0.002, and <0.001. These results indicate that the state-conditioned representation is particularly beneficial for the complex real-robot-informed setting.

Q3: The modifications used to adapt prior methods need to be described more clearly

We expanded Appendix D.3 and moved the complete adaptation summary into Table 9. Every method is evaluated without current or past target-field observations, and all methods receive the same sensor information.

MethodOriginal inputAdaptation used in our evaluation
DeepONetInput function and query coordinatesSensor embedding in the branch and node coordinates in the trunk
S-DeepONetSequential loading historyNative recurrent branch with sensor history and coordinate trunk
SHREDSparse sensor historyNative recurrent encoder with a training-set POD basis
MeshGraphNets-adaptedPrevious full fieldField channels removed and sensor embedding broadcast to nodes
BSMS-GNN-adaptedPrevious full fieldSame sensor conditioning as MeshGraphNets with multi-scale processing retained
LSTM-GNN-adaptedSensor history and meshNative recurrent conditioning retained and constitutive physics losses removed
GNOTFunction and geometry queriesSensor embedding supplied through global conditioning slots
TransolverField or function samples on pointsCoordinates concatenated with sensor embedding
ErwinPoint-set field featuresSensor embedding concatenated to point features

The accompanying paragraphs specify the implementation basis and preserved components. For example, MeshGraphNets retains its encode-process-decode structure, edge features, and message-passing depth, but the prohibited field input is replaced by the broadcast sensor embedding. SHRED retains its native recurrent encoder and frozen POD decoder, with the POD rank tuned on the validation split. LSTM-GNN retains recurrent observation conditioning and mesh message passing, while its constitutive physics losses are removed because they are specific to the original setting.

All baselines are tuned on the validation split using the same training objective and data splits. The search includes hidden width, depth, learning rate, and method-specific structural parameters. The sensor embedding dimension is fixed at 128 for all methods. We therefore compare the strongest validated task-adapted configurations we obtained, while explicitly avoiding claims about the original methods in their native input settings.

Q4: How does the method perform when using larger datasets?

We expanded the mechanical-leg evaluation from 500 independent conditions in the previous synthetic Go1Leg dataset to 4,500 independent conditions in Go1Leg-Real. This is a ninefold increase in independent operating conditions, rather than an increase obtained by sampling additional overlapping windows from the same trajectories.

Each Go1Leg-Real condition contains 300 time steps of physical robot sensor measurements. The dataset uses 3,375 training, 562 validation, and 563 test conditions, with all normalization statistics computed on the training split. The mesh contains 22,515 nodes and 93,983 elements, and each sample requires prediction of a six-component dynamic stress field from a 50-step sensor window.

Under this larger and more complex setting, SONAR obtains:

MetricSONARStrongest baselineRelative reductionpp-value
NMSE0.1673 ± 0.00530.2019 ± 0.006217.1%<0.001
RelL20.3129 ± 0.00950.3458 ± 0.01279.5%0.002
VM-MaxRE9.8 ± 0.4914.2 ± 0.5731.0%<0.001

The complete ten-method comparison is provided under Q2. These results show that SONAR achieves a consistent and substantial advantage on the larger real-robot-informed dataset. Because the new benchmark also changes the sensor source and label construction, we present this as evidence on the expanded benchmark rather than as a controlled data-scaling study.

Reviewer_zzFuOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SONAR, a sensor-conditioned neural operator for predicting next-step full-field on unstructured meshes from short windows of heterogeneous observed state sequences. The method combines adaptive spectral rotation, Gram whitening, and curvature-guided residual correction to capture global field distributions and localized high-frequency errors. Experiments on three physical benchmarks elastic-beam, heat-plate, and mechanical-leg benchmarks demonstrate that SONAR achieves the best accuracy among the compared methods.

Critical strengths

1. The distinction between full-field autoregressive simulation, field reconstruction, and field inference from heterogeneous system-level measurements is valuable.

2. Ablation study supports improvement over each architecture component.

Critical weaknesses

1. The paper is very difficult to follow, and the novelty of the proposed method is not clear.

The method uses several existing components, such as GRU, message passing, Gram whitening, and a curvature-guided residual branch. However, it is not clear what is fundamentally new in the proposed approach compared to existing sensor-conditioned graph neural networks, apart from combining these components.

2. The experimental evaluation is limited to fixed geometries.

In Section 4.1, the authors state that “Within each dataset, geometry and materials are fixed and the observed state sequence is the only per-sample variation.” Although the authors acknowledge this limitation in Section 5, all experiments are conducted on a single fixed mesh for each dataset. Since the proposed method focuses on full-field prediction on unstructured meshes, evaluating its performance on different mesh resolutions or unseen geometries would provide stronger evidence of its generalization ability.

Questions for authors

Q1: How well does SONAR generalize to unseen geometries or different mesh resolutions, given that all experiments use fixed meshes?

Q2: How much computational cost does Gram whitening introduce during training and inference?

Q3: Could you compare SONAR with a learned but sensor-independent graph basis to better understand the benefit of adaptive spectral rotation?

Q4: Some improvements over the strongest baselines are relatively small. How meaningful are these improvements.

3 comments

We sincerely thank the reviewer for the careful assessment and for identifying the main barriers to evaluating our work: unclear presentation, insufficiently isolated novelty, fixed-mesh evaluation, and the practical meaning of relatively small improvements. We have substantially rewritten the manuscript and added experiments that directly address each concern. In particular, the revised paper now presents state-conditioned basis generation as the central contribution, includes a learned sensor-independent basis control, evaluates cross-resolution and unseen-geometry generalization, reports whitening and end-to-end computational costs, and evaluates SONAR on the larger Go1Leg-Real benchmark with physical robot sensor inputs.

Critical Weakness 1: Presentation and novelty

We apologize that the initial presentation obscured the methodological contribution. The revised manuscript reorganizes the introduction and method around one explicit distinction:

  • A fixed-basis decoder uses a shared spatial basis for all samples and changes only the basis coefficients.
  • A direct sensor-conditioned mesh decoder changes node features but does not explicitly separate spatial subspace selection from field synthesis.
  • SONAR generates a new mesh basis for each observation window and predicts coefficients within that sample-specific basis.

The central operation is now stated directly in Section 3.2, Eq. 2:

\qquad \hat{Y}=\hat{Y}_{\mathrm{glob}}+\hat{Y}_{\mathrm{res}}.$$ Thus, the observation history controls both the spatial subspace and the coordinates within that subspace. The novelty is not presented as the individual use of a GRU, message passing, whitening, or residual correction. It is the state-conditioned basis construction and the explicit separation of per-sample subspace selection from coefficient regression. Gram whitening gives the generated basis a common scale across samples, while the localized residual branch represents structure outside the global rank-$R$ approximation. We also clarified the terminology. Adaptive spectral rotation refers to changes in learned basis directions across observation states, not to precomputed eigenvectors of a fixed mesh operator. The active node set is a computational subset for residual correction and is unrelated to active learning. Detailed proofs are consolidated in Appendix A, while the main paper retains only their design implications. Most importantly, the new learned shared-basis control isolates the proposed mechanism experimentally. It retains the generator, coefficient head, rank, whitening, residual branch, and training protocol, but removes sensor input to the basis generator. As shown in Table 4, this learned but sensor-independent basis is consistently worse than the conditioned basis. We therefore believe that the revised manuscript now identifies and tests a specific representational contribution rather than presenting a combination of existing modules. ## Critical Weakness 2: Evaluation on fixed geometries We agree that the original experiments did not adequately test the claimed mesh generalization. The revised Section 4.6 and Table 7 now include direct cross-resolution and unseen-geometry evaluations on ElasticBeam-Syn. These tests use test discretizations or geometries that are absent from training; retraining on a new mesh is not counted as cross-mesh generalization. | Evaluation | SONAR NMSE | MeshGraphNets-adapted NMSE | Learned shared-basis NMSE | | ----------------- | ------------------: | -------------------------: | ------------------------: | | Coarser mesh | **0.0420 ± 0.0073** | 0.0850 ± 0.0081 | 0.0960 ± 0.0068 | | Finer mesh | **0.0580 ± 0.0095** | 0.1100 ± 0.0079 | 0.1400 ± 0.0107 | | Unseen geometries | **0.1200 ± 0.0108** | 0.2100 ± 0.0093 | 0.2500 ± 0.0132 | SONAR achieves the lowest error in all three evaluations. The error increase from 0.0420 on the coarser mesh to 0.1200 on unseen geometries also makes the remaining difficulty explicit. We have narrowed the conclusion accordingly: these results demonstrate improved generalization in the tested settings, not generalization to arbitrary physical systems or mesh families.

Q1: How well does SONAR generalize to unseen geometries or different mesh resolutions?

We added the direct evaluations reported in Table 7. For cross-resolution tests, the model is trained on the original ElasticBeam-Syn mesh and evaluated on either a coarser or finer mesh. For unseen-geometry evaluation, training and testing use geometry-disjoint sets.

SONAR obtains NMSE values of 0.0420 ± 0.0073, 0.0580 ± 0.0095, and 0.1200 ± 0.0108 on the coarser mesh, finer mesh, and held-out geometries, respectively. The corresponding MeshGraphNets-adapted results are 0.0850 ± 0.0081, 0.1100 ± 0.0079, and 0.2100 ± 0.0093, while the learned shared basis obtains 0.0960 ± 0.0068, 0.1400 ± 0.0107, and 0.2500 ± 0.0132.

These results support a benefit of the state-conditioned representation under the evaluated resolution and geometry shifts. We also report the increased error on unseen geometries rather than presenting the tests as evidence of unrestricted geometric generalization.

Q2: How much computational cost does Gram whitening introduce during training and inference?

We added a dedicated whitening profile in Appendix D.6, Table 10. The measurements are conducted on Go1Leg-Real using the same hardware, precision, and batch size. Forward and backward times isolate the normalization operation, and the final column reports full-model batch-one latency.

Basis rank RRWhitening forwardWhitening backwardPeak memoryFull-model batch-one latency
160.18 ms0.42 ms42 MB118 ms
320.35 ms0.80 ms78 MB142 ms
640.85 ms1.90 ms182 MB185 ms
1282.40 ms5.80 ms564 MB265 ms

At the default rank R=64R=64, whitening takes 0.85 ms in the forward pass and 1.90 ms in the backward pass, with 182 MB peak memory. Increasing RR from 16 to 128 raises forward whitening time from 0.18 ms to 2.40 ms and full-model latency from 118 ms to 265 ms. These results make the cost of normalization explicit rather than relying only on asymptotic complexity.

The revised Table 6 additionally reports end-to-end parameters, training GPU hours, training and inference memory, and inference latency for SONAR and representative baselines.

Q3: Could you compare SONAR with a learned but sensor-independent graph basis?

Yes. We added precisely this control in Section 4.4, Table 4, with implementation details in Appendix D.5. The learned shared graph basis is trained end to end but receives no sensor input to its basis generator. Its coefficient head remains sensor conditioned, and its rank, whitening, residual branch, training objective, and evaluation protocol are unchanged. This isolates observation dependence from the more general benefit of learning a basis.

Basis representationSensor input to basisBasis varies by sampleElasticBeam-Syn NMSEGo1Leg-Real NMSE
Fixed PODNoNo0.0612 ± 0.00340.3027 ± 0.0126
Learned shared graph basisNoNo0.0435 ± 0.00380.2989 ± 0.0082
SONAR conditioned basisYesYes0.0249 ± 0.00090.1673 ± 0.0053

Learning a shared basis improves over fixed POD, but conditioning the basis on the observation state provides a further substantial reduction: from 0.0435 to 0.0249 on ElasticBeam-Syn and from 0.2989 to 0.1673 on Go1Leg-Real. These correspond to relative NMSE reductions of approximately 42.8% and 44.0% over the learned sensor-independent basis. This control provides direct evidence for the benefit of adaptive basis generation.

Q4: Some improvements over the strongest baselines are relatively small. How meaningful are these improvements?

We agree that the practical interpretation should distinguish effect size, training variability, and the evaluation setting. The revised manuscript reports five-seed results and two-sided Welch tests against the strongest baseline for every dataset and metric. It also explicitly identifies the comparisons that do not support a statistically confirmed improvement.

DatasetMetricSONARStrongest baselineRelative reductionpp-value
ElasticBeam-SynNMSE0.02490.02635.3%0.043
ElasticBeam-SynRelL20.16000.16060.4%0.872
ElasticBeam-SynVM-MaxRE2.7302.8514.2%0.037
HeatPlate-SynNMSE0.00050.000728.6%0.004
HeatPlate-SynRelL20.02810.031811.6%<0.001
HeatPlate-SynVM-MaxRE0.7340.7421.1%0.054
Go1Leg-RealNMSE0.16730.201917.1%<0.001
Go1Leg-RealRelL20.31290.34589.5%0.002
Go1Leg-RealVM-MaxRE9.814.231.0%<0.001

SONAR has the best mean in all nine comparisons, and seven comparisons satisfy p<0.05p<0.05 without multiplicity adjustment. We do not claim that every improvement is statistically significant. In particular, ElasticBeam-Syn RelL2 differs by only 0.4% with p=0.872p=0.872, so the two leading methods are statistically tied on this metric. HeatPlate-Syn VM-MaxRE differs by 1.1% with p=0.054p=0.054, which we describe as marginal evidence rather than a confirmed improvement. The tests quantify seed-level training variability, not uncertainty from drawing a new test set.

The practical advantage is clearer on the new Go1Leg-Real benchmark. This dataset contains 4,500 independent operating conditions with physical robot sensor histories and a complex 22,515-node leg mesh. SONAR reduces NMSE by 17.1%, RelL2 by 9.5%, and VM-MaxRE by 31.0% relative to the strongest baseline for each metric, with pp-values of <0.001, 0.002, and <0.001, respectively. Thus, although some gains on the smaller synthetic benchmarks are modest, the advantage becomes more pronounced on the larger and more complex real-robot-informed setting. For industrial and embodied applications that require field reconstruction from available sensor histories, these reductions represent higher global field fidelity and lower localized tail error.

Figure 3 complements the quantitative results with predicted fields and error maps across methods. The visual comparison shows that SONAR better preserves localized field structures and reduces spatially distributed errors, consistent with its lower global and tail-error metrics. We use this figure as qualitative evidence and rely on the numerical tables for the quantitative conclusions.

We have also added computational measurements so that these accuracy gains can be weighed against their cost. We therefore do not argue that SONAR dominates every method in every practical respect. The revised claim is narrower: SONAR provides consistent accuracy benefits under the sensor-only contract, with especially clear improvements on Go1Leg-Real, while its additional computational cost is explicitly reported.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.