acceptodds
Under review as a conference paper at ICLR 2027

EigenRR: Learning Tetrahedral Search Subspaces for Impact Sound Synthesis

Abstract

Synthesizing a realistic sound for a struck object first requires computing how that object vibrates. Physically, the object is meshed into many small tetrahedra (a finite-element discretization), and a very large eigenvalue problem is then solved. Classical solvers such as ARPACK and LOBPCG are accurate but too slow. Learning-based approaches either regress modal quantities directly or learn a search subspace on a voxelized mesh that requires interpolation to tetrahedral FEM meshes, risking transfer error. EigenRR learns an overcomplete search subspace on the tetrahedral FEM mesh, and Rayleigh–Ritz extracts approximate eigenpairs using that mesh's assembled stiffness and mass matrices. The resulting modes are consistent with the physical operator, so they also warm-start an iterative solver. Twenty LOBPCG steps from this prediction reach a mel-frequency error of , and more accurate than NeuralSound and random initializations under the same budget. One forward pass already beats 20 iterations from a random start. We build VibraVerse++, a geometry–acoustics dataset of tetrahedral meshes. We train and evaluate on it, and release the meshes together with the assembled FEM matrices and reference eigenspectra. The method returns accurate frequencies and mode shapes at interactive speed.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
audio, modal analysis, Rayleigh–Ritz, sound synthesis, dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_r5enOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the modal impact-sound synthesis, concerning the cost of solving the FEM generalized eigenproblem. The proposed EigenRR trains a graph neural net on the tetrahedral mesh to predict an overcomplete 80-direction search subspace. A differentiable Rayleigh–Ritz layer, using the assembled stiffness and mass matrices, extracts the first 40 elastic modes. These Ritz vectors can also warm-start LOBPCG. The method is trained with subspace, mel-frequency and basis-regularization losses on a newly generated VibraVerse++ geometry-acoustics dataset. On the test set, the forward pass (39 ms) and 20 warm-started LOBPCG steps (0.19 s) give lower mel error than NeuralSound, direct regression and random initialization. Ablations support the Rayleigh–Ritz layer and the loss terms. Qualitative results include spectrograms, location-dependent synthesis and strike-point estimation.

Critical strengths
  1. Clear motivation on the proposed approach. The decomposition is principled and well motivated. Learning only the search subspace and leaving eigenpair extraction to Rayleigh–Ritz on the target FEM operator is an elegant idea. Parameterizing the subspace with a graph network directly on the tetrahedral mesh avoids the voxel-to-mesh transfer error of prior work. It also yields mode shapes, which enables location-dependent synthesis that eigenvalue-only regression cannot provide.
  2. Quantitative results. As the authors claim, the accuracy–latency trade-off improves consistently over NeuralSound and random initialization at every LOBPCG cap tested (0–100 steps, Fig. 2/Table 1). The Rayleigh–Ritz ablation (Table 2) supports the design choice. Release of meshes, assembled matrices and reference spectra is a useful community resource.
Critical weaknesses
  • W1. Practical significance. of the gains is not yet established. Improvements are reported as multiplicative factors in mel MSE (e.g., 2.4E-4 vs. 3.7E-4), but it is unclear whether such differences are audibly significant. The synthesis evidence (Fig. 3) is limited to two samples, and differences between EigenRR and NeuralSound are hard to see; for the wood example spectrogram, since the damping is strong most of the frames remain silent. The text itself notes these samples are not a test-set metric. There is no listening study or perceptually grounded error unit (e.g., cents), and no variance across seeds, confidence intervals, or per-sample distributions. Standard acoustic analyses would make the benefit much clearer: band-wise T60/EDT (decay is a deterministic function of the modal frequencies here, so this directly tests error propagation), overlaid magnitude spectra, and spectrogram difference plots. The manuscript is short, and reporting in several places is thin.
  • W2. Positioning to the prior works. The positioning of the proposed method, relative to prior work, seems a little incomplete. Learning a subspace followed by Rayleigh–Ritz and/or LOBPCG warm-starting has been explored in neural eigensolver work not discussed here, including [N1], [N2], [N3], [N4]. I believe the last [N4] also predicts a redundant basis refined by Rayleigh–Ritz, yet this manuscript could have been submitted before [N4] come out. But, the contribution could be framed as adapting these paradigm to tetrahedral FEM modal analysis for sound. It would be nicer to clarify the positioning of this work relative to these prior works explicitly.

[N1] Fanaskov, V., Trifonov, V., Rudikov, A., Muravleva, E., & Oseledets, I. (2026, April). Deep learning for subspace regression. In International Conference on Learning Representations (Vol. 2026, pp. 115208-115239).

[N2] Dai, X., Fan, Y., & Sheng, Z. (2024). Subspace method based on neural networks for eigenvalue problems. arXiv preprint arXiv:2410.13358.

[N3] Li HA, Sun JI, Zhang ZH. Deep Eigenspace Network for Parametric Non-self-adjoint Eigenvalue Problems. arXiv preprint arXiv:2512.20058. 2025 Dec 23.

[N4] Yang, Z., Du, T., & Liu, L. (2026). Learning Laplacian Eigenspace with Mass-Aware Neural Operators on Point Clouds. arXiv preprint arXiv:2605.24390.

Questions for authors

Clarifications:

  1. Are the reported speedups consistent across repeated runs with seed change?
  2. How were the glass and wood examples selected? Do the reported advantages hold over many test objects?
1 comment

Thank you for recognizing the advantages of combining learning with numerical methods and of computing mode shapes directly on tetrahedral meshes. Below, we clarify the evidence supporting the consistency of the reported gains, the scope of our evaluation, and the selection of the synthesis examples. We also illustrate how modal frequency errors translate into pitch differences to explain the acoustic relevance of the reported metric.

Q1. Consistency of speedups across repeated runs

We did not vary the random seed during testing. However, the following two considerations support the consistency of the observed gains. First, for a fixed model and FEM system, seed selection does not change the network architecture or the nominal computational cost per solver iteration. Second, the separately executed experiments in Sections 5.2 and 5.3 provide empirical evidence for this consistency. As noted in Section 5.3, GPU floating-point nondeterminism, numerical variability in LOBPCG, and differences in machine load produced different accuracy and timing measurements from Table 1. Despite these variations, EigenRR maintained its favorable accuracy–latency trade-off in both experiments. This consistency supports the robustness of the observed advantage to execution variability, while not constituting a controlled multi-seed evaluation.

Q2. Example selection and performance across the test set

Example selection. The examples were selected simply according to their order in the dataset: the glass and wood examples correspond to the second and third samples, respectively, while the first sample was used for the experiment in Section 5.7. These examples serve as qualitative illustrations of the synthesized sounds and their spectral characteristics for two different materials.

Performance across the test set. The quantitative results in Sections 5.2 and 5.3 are averaged over the entire test set and show an improved accuracy–latency trade-off for EigenRR. The reported advantages therefore reflect aggregate performance across test objects and are not limited to the two qualitative examples. These results support an average advantage under the evaluated settings; they do not imply that EigenRR outperforms every baseline on every individual object.

Effects of frequency deviations on perceived sound

Errors in modal frequencies shift the resonance peaks of synthesized sounds and, under our damping model, also affect their decay rates. Errors in mode shapes alter the relative amplitudes excited at a given strike location. We use the following example to provide an intuitive, qualitative illustration of how frequency deviations affect perceived pitch.

Consider a hypothetical glass object with a fundamental resonance at C4 (261.63 Hz). Suppose that, after a forward pass and 20 LOBPCG iterations, EigenRR predicts D4 (293.66 Hz), while NeuralSound predicts D♯4 (311.13 Hz). Under our normalization, the corresponding single-frequency mel MSE values are approximately 9.52×10−59.52\times10^{-5} and 2.23×10−42.23\times10^{-4}, respectively. These small numerical values correspond to pitch deviations of 200 and 300 cents, with the two predictions separated by a full semitone. This example illustrates that a small absolute difference in normalized mel MSE can correspond to a substantial musical interval, highlighting why accurate modal frequencies matter for sound synthesis.

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

The paper describes a method for predicting the sound of a struck object based on a combination of a learned aproach from data followed by classical solver steps. Rather than directly predicting the frequencies from an input object defined by its tetrahedral FEM mesh, it learns the (overcomplete) search space that functions as a warm start for classical iterative methods to get the final result. This speeds up the classical method (one forward pass beating 20 iterations of the iterative method) and leads to better predictions at similar budgets compared to other existing work.

Critical strengths

The paper seems a highly technical contribution that succesfully combines numerical mathematical techniques with AI. I believe this direction of still exploiting existing numerical approaches in AI for science, rather than plain black-box learning, is very valuable, and appreciate the elegance in which they combine in this paper.

The VibraVerse++ dataset can be very useful for improving research in this field.

Critical weaknesses

It is unclear to me what this paper exactly adds that is different from existing work like NeuralSound, which also combines a learning based approach with LOBPCG and Rayleigh-ritz. This could also imply my reading is wrong, in which case I would suggest the paper could be improved by clarifying this further.

One big difference seems to relate to the mesh being tetrahedral instead of the voxel grid, which consequently implies the graph model being used in the first place. Here I do believe the different grids/mesh types leads to the risk of an unfair/unclear comparison; if the tetrahedral mesh is just a better one, i.e. more closely aligning with the real object, part of the advantages of the authors method over NeuralSound could come from that, and I think it would be fair to have ablations to clearly show the distinction. Would a NeuralSound-type method also work on the grids used in EigenRR, and vice-versa? What's more, the argued disadvantage of having to first interpolate between grids (incurring both extra compute costs as well as interpolation errors) seems a bit constructed, as it implies already having a mesh A and model B mismatch, which could equally well be mesh B and model A.

Questions for authors

I found the tone occasionally muddying the distinction between the authors contribution and existing methods/ideas. As technical implementation details and generic background knowledge are blended with precise method decisions made by the authors, it is unclear to a less knowledgeable reader what the novelty is (to what extent is X being suggested and explicitly tested by this paper, vs a canonical technique that others also apply in some form). It would help me a lot with my review if the authors could clarify for their method sections which parts are (and to what extent) novel contributions.

1 comment

Thank you for appreciating the combination of learning with numerical computation. Your uncertainty about what is new is understandable: the introduction and contribution list do not consistently separate established ingredients from our specific implementation. We will correct that presentation.

Which parts of the method are new

The intended distinction, section by section, is:

SectionEstablished ingredientsWhat EigenRR specifies or evaluates
4.1 OverviewLearning a useful search space for eigensolvingThe tetrahedral modal-analysis pipeline and output interface
4.2 Neural Subspace PredictionGraph U-Nets, positional encoding, and residual MLPsBuilding graph structure on tetrahedral meshes, and configure those modules to predict vector-valued trial directions at tetrahedral vertices
4.3 Differentiable Rayleigh-Ritz LayerMass orthogonalization and projected eigenpair extractionApplication to the target assembled K/M, with documented stabilization and extraction choices
4.4 Training ObjectiveThe standard mel-frequency scale and its frequency-to-mel mappingOur mass-weighted subspace alignment loss, normalized mel-frequency loss, and log mass-energy basis regularizer, combined to improve modal and frequency accuracy while stabilizing basis scaling
4.5 Warm-Started LOBPCG RefinementLearned warm-starting of iterative solversInitialization and evaluation on the tetrahedral system

Sections 4.3 and 4.5 draw on established methods, including those used in NeuralSound: extracting eigenpairs through mass orthogonalization and Rayleigh–Ritz projection, and warm-starting LOBPCG, respectively. Our contributions include a graph-based modal solver that learns search subspaces directly on tetrahedral FEM meshes, a joint training objective for subspace alignment, perceptual frequency accuracy, and basis stabilization, and the VibraVerse++ operator-level dataset with a systematic evaluation of accuracy and efficiency.

Why compare on a tetrahedral target

Our problem is to approximate the modal solution of a specified tetrahedral FEM system, so comparing all refined solutions against its reference spectrum measures how quickly each initialization pipeline approaches the desired discrete solution. Because a method native to the target representation avoids the adaptation cost incurred by a method using a different representation, this benchmark evaluates the complete pipelines on a tetrahedral target. Its conclusions are therefore specific to this setting; a voxel-target application could favor a voxel-native method.

We also emphasize the inherent geometric flexibility of tetrahedral meshes compared with regular voxel-based hexahedral grids. Their boundary conformity and local refinement better preserve curved surfaces and thin structures, avoiding the staircase artifacts introduced by voxelization. Because geometric distortions affect both mass and stiffness, this greater fidelity can, with adequate mesh quality and resolution, reduce geometry-induced spectral errors and yield modal solutions closer to the underlying continuum vibration problem.

Can the methods operate on each other's grids

The underlying learned-subspace/Rayleigh–Ritz idea can be used on either discretization. NeuralSound's specific sparse convolutional implementation assumes voxel structure, so operating directly on an unstructured tetrahedral mesh requires changing that backbone or an adaptation layer. Conversely, a graph predictor can in principle operate on the vertex graph of a voxel FEM discretization, with the corresponding degrees of freedom and K/M.

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

EigenRR targets impact sound synthesis: a graph neural network predicts an overcomplete search subspace on tetrahedral FEM meshes, and a differentiable Rayleigh–Ritz layer projects and solves the problem to produce the first 40 elastic modes, which can also warm-start LOBPCG. Its algorithmic contributions are a tetrahedral-mesh-native GNN instantiation (eliminating voxelization transfer error) and three training losses—subspace alignment, mel-frequency, and basis regularization. On VibraVerse++, it reaches a mel-frequency error of 7.2×10⁻³ in 39 ms, about 2.1× better than NeuralSound; the error can be significantly reduced after refinement with LOGPCG.

Critical strengths

This paper trains tetrahedral-mesh-native GNN to predict the search subspace, yielding non-trivial gains in both accuracy and efficiency. The zero-step forward pass takes only 39 ms (3.8× faster than NeuralSound's 148 ms) and attains a mel-frequency error of 7.2×10⁻³, about half of NeuralSound's. After 20 LOBPCG iterations it runs in 0.19 s with an error of 2.4×10⁻⁴; this is 1.55× (3.7×10⁻⁴/2.4×10⁻⁴)and 63× (1.5×10⁻²/ 2.4×10⁻⁴) better than NeuralSound and random initialization, respectively. Random initialization still cannot reach this accuracy after 100 iterations (~1.05 s, 7.5×10⁻⁴), corresponding to roughly a 5× reduction in refinement time.

Critical weaknesses
  1. The differentiable Rayleigh–Ritz layer—SVQB orthonormalization, K/M projection, the small projected eigenproblem—is mathematically similar to NeuralSound (TOG 2022), whose Eq. (5) already states the same projection and SVQB. The first contribution should be narrowed to the tetrahedral-mesh-native GNN instantiation and new training losses, with a component-wise comparison.

  2. The voxelization-transfer-error claim is confounded with the backbone change (sparse U-Net → GNN), with no voxel-vs-tetrahedron control on the same meshes. Key hyperparameters are also unablated: only the LOBPCG cap is swept, while the subspace size s=80 and the 1e-8 threshold are fixed, so gains cannot be attributed to representation, architecture, or subspace size.

Questions for authors

VibraVerse++ is produced entirely through a single AI asset-generation pipeline, namely "text-to-image → image-to-3D → tetrahedralization with fTetWild," and covers only 10 homogeneous materials, with 64 ARPACK reference eigenpairs per object. Given the single data source and the absence of real scanned or engineering CAD meshes, the proposed algorithm should ideally be evaluated on real scanned data. Is it possible to report the testing results on CAD objects?

2 comments

Thank you for recognizing EigenRR's improvements in accuracy and efficiency. We understand your concern about generalization beyond the synthetic data-generation pipeline used for VibraVerse++, and have conducted additional experiments on CAD objects to address this concern.

Evaluation on CAD objects

Experimental setup. We sampled 300 CAD models from the ABC dataset (Koch et al., CVPR 2019) and randomly assigned each model one of the ten materials used in VibraVerse++. Applying the same preprocessing and filtering procedures as for VibraVerse++ yielded 90 samples suitable for linear tetrahedral finite-element simulation. We directly evaluated the model trained on VibraVerse++ on these CAD objects, without retraining or fine-tuning.

Results. We compare the same six method configurations as in the main experiments (Sec. 5.1), reporting subspace error, mel-frequency MSE, and per-sample runtime.

MethodSubspace error ↓Mel-frequency MSE ↓Time (s) ↓
VibraVerse (direct regression)—0.1084610.063
NeuralSound0.12050.0057500.078
EigenRR (forward only)0.10460.0040630.042
Random init. + LOBPCG200.26430.0057520.146
NeuralSound + LOBPCG200.06350.0000950.247
EigenRR + LOBPCG200.05850.0000870.190

Conclusion. EigenRR's performance on CAD objects remains consistent with the main experiments: it is faster and achieves lower subspace error and mel-frequency MSE than baseline methods in both the forward-only and LOBPCG20 settings. Across all tested configurations, forward-only EigenRR is the fastest, while EigenRR + LOBPCG20 achieves the lowest reported subspace error and mel-frequency MSE. We will include this experiment in the revised manuscript.

Contributions relative to NeuralSound

Our contributions comprise a graph-based modal solver operating directly on tetrahedral FEM meshes, a joint supervised training objective, and the accompanying operator-level dataset. The distinction from NeuralSound is summarized below.

ComponentNeuralSoundEigenRR
Subspace predictionSparse 3D U-Net on voxel hexahedral meshesGraph network predicting an overcomplete basis directly on tetrahedral meshes
Training objectiveEigenpair residual lossMass-weighted subspace alignment, normalized mel-frequency loss, and log mass-energy basis regularization
Numerical extraction and refinementMass orthogonalization, Rayleigh–Ritz, and LOBPCGThese numerical ingredients applied to the target tetrahedral FEM operators

Direct prediction on the target mesh removes the need for voxel-to-mesh interpolation of mode shapes.

Thanks for adding the experiments on CAD objects. What I am really concerned is that the CAD objects contain sharp or very small--scale features, which might influcence the simulaiton results. Thus, this concern are partially addressed. While ground-truth results are not available to evaluate the advantage of EigenRR over NerualSound, it is still valuable to see how EigenRR works against NeuralSound on CAD objects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.