- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a method called TIDAL to reconstruct 3D complex ocean acoustic fields from sparse H/V observations. The main idea is to use wave physics to derive a spatial-angular tensor representation (S/c) representation. The authors further connect this low-rank form to LL1 tensor decomposition and give recovery conditions based on existing results. Based on this representation, they design a neural network with attention and a structured decoder. Experiments on simulated ocean acoustic fields show better reconstruction accuracy than several baselines.
Critical strengths
-
The problem is meaningful. Reconstructing complex acoustic fields from sparse measurements has practical value.
-
The paper is generally well organized. The motivation and the overall framework are easy to follow.
-
The experimental results show improvements over the selected baselines.
Critical weaknesses
-
Novelty is limited. The model applies existing tensor decomposition and neural network representation to a standard tensor recovery application.
-
The recovery theorem is based on an exact low-rank LL1 decomposition. However, the real TIDAL model uses dense spatial components. This means the theoretical guarantee does not apply to the proposed model. The theory is also based on existing results.
-
It is unclear whether the sampling conditions in the theory are satisfied in the actual experiments. As a result, I do not think the theoretical part provides a strong justification for the actual neural network. The theory and the implementation seem to be two somewhat separate parts.
-
The proposed neural network learns both spatial and angular components freely. The angular factors are not fixed Fourier basis. So the actual decoder looks more like a general learnable tensor factorization. I am not convinced that the proposed representation contains a physics-specific insight.
-
All experiments are conducted on simulated acoustic fields generated from a limited set of configurations. There is no convincing experiment on unseen ocean real environments. The proposed model uses information from multiple angular directions, while several baseline reconstruct individual angular slices. This is an unfair setting. Also, some baselines are adapted from quite different tasks. For example, MxT is designed for image inpainting only.
Questions for authors
My questions are listed in Critical Weaknesses.
4 comments
Q5. 3D baselines and generalization
To provide direct comparisons with methods that jointly exploit all 72 angular directions, we introduced four additional 3D reconstruction baselines. These include two training-free methods, PRGD-Tucker [1] and ShearLab 3D [2], which perform three-dimensional low-rank tensor completion and shearlet-based sparse reconstruction, respectively. We also included two training-based generative methods, conditional 3D EDM and conditional 3D GAN [3]. All four methods reconstruct the complete 200 × 200 × 72 acoustic field, jointly utilizing the spatial and angular dimensions. We evaluated these methods under the same H/V sampling geometries as TIDAL. The results for the primary H15/V15 setting are summarized below.
Table 1. Reconstruction performance under H15/V15 sampling (Unobserved relative RMSE, mean ± std).
| Method | Equispaced | Shifted | Jittered |
|---|---|---|---|
| PRGD-Tucker [1] | 1.034 ± 0.058 | 1.009 ± 0.056 | 1.028 ± 0.064 |
| ShearLab 3D [2] | 0.972 ± 0.038 | 0.975 ± 0.013 | 0.978 ± 0.012 |
| Conditional 3D EDM [3] | 0.356 ± 0.071 | 0.358 ± 0.071 | 0.372 ± 0.075 |
| Conditional 3D GAN [3] | 0.544 ± 0.087 | 0.543 ± 0.082 | 0.559 ± 0.081 |
| TIDAL | 0.269 ± 0.085 | 0.274 ± 0.085 | 0.295 ± 0.089 |
We additionally examined H10/V10 and H20/V20 sampling to assess performance across different observation densities.
Table 2. Average unobserved relative RMSE across the three sampling geometries.
| Method | H10/V10 | H15/V15 | H20/V20 |
|---|---|---|---|
| PRGD-Tucker [1] | 1.1180 | 1.0234 | 0.9627 |
| ShearLab 3D [2] | 0.9916 | 0.9751 | 0.9557 |
| Conditional 3D EDM [3] | 0.4101 | 0.3620 | 0.3376 |
| Conditional 3D GAN [3] | 0.6200 | 0.5487 | 0.5216 |
| TIDAL | 0.3370 | 0.2795 | 0.2620 |
These results further demonstrate the effectiveness of TIDAL in comparison with methods that jointly model the complete 72-angle acoustic field.
Regarding generalization, our dataset is generated using the KRAKEN solver with realistic oceanographic and bathymetric information, spanning multiple source locations, frequencies, source depths, and propagation conditions. The simulation settings and evaluation scope are described in Sec. 6.1, while generalization to measured acoustic fields and unseen ocean environments is discussed as future work in Sec. 7. Additional reconstruction results are now reported in Table 1 and Appendix C (Table 3).
[1] Zhang Y, Bian F, Zhang X, et al. Preconditioned Riemannian gradient descent algorithm for low-multilinear-rank tensor completion[C]//Forty-second International Conference on Machine Learning. 2025.
[2] Kutyniok G, Lim W Q, Reisenhofer R. Shearlab 3D: Faithful digital shearlet transforms based on compactly supported shearlets[J]. ACM Transactions on Mathematical Software (TOMS), 2016, 42(1): 1-42.
[3] Oommen V, Khodakarami S, Bora A, et al. Learning turbulent flows with generative models for super resolution and sparse flow reconstruction[J]. Nature Communications, 2026, 17(1): 3707.
This reviewer appreciates the new experiments and clarifications and would like to increase the score to 2. However, this reviewer believes that the overall novelty and new conceptualization remain limited in this manuscript. It is essentially an application of some existing tensor models to a new type of data, with considerable engineering effort. I think a score of 2 is fair.
We thank the reviewer for reconsidering our work. We would like to further clarify the central conceptual contribution of this work, which may not have been sufficiently emphasized in our previous response.
Our approach does not start by applying an existing tensor decomposition to acoustic data. Instead, we begin with the governing wave physics and derive a structured tensor representation that reveals the underlying spatial–angular structure of acoustic fields. This physics-derived representation then serves as a principled inductive bias for neural reconstruction, directly guiding how the neural network organizes and reconstructs the acoustic field. In this way, the architecture is designed based on the physical structure revealed by the governing equations, providing a way to incorporate wave physics into neural reconstruction without requiring known environment-dependent physical parameters.
Therefore, our main contribution lies in establishing the connection from wave physics to structured tensor representation and then to neural architecture design, rather than simply applying existing tensor models to a new type of data.
We thank the reviewer for the comments and the opportunity to clarify our contributions and theoretical motivation.
Our main contribution lies in deriving a structured representation from wave propagation physics and incorporating it into neural reconstruction under sparse H/V observations. The LL1 representation was not obvious a priori; it was revealed by our derivation from the governing physics equations.
Specifically, the physics-derived S/c representation guides the design of the structured decoder, while the H/V sampling geometry motivates the sampling-aligned axial attention. These two components jointly adapt the reconstruction architecture to the physical field structure and observation geometry.
The ablation studies in the original paper further demonstrate the effectiveness of this design. Compared with direct-output decoding, the S/c decoder consistently improves reconstruction accuracy, with particularly clear advantages under limited training data and reduced model capacity. These results support the value of the proposed representation-guided reconstruction framework.
The theoretical analysis addresses a fundamental question in practical ocean sensing: whether structured H/V observations can uniquely determine the full 3D acoustic field. Starting from the physics-derived S/c representation, we consider its low-rank LL1 specialization and build on established tensor recovery results to characterize sufficient recoverability conditions for complex acoustic fields under H/V sampling. This establishes a principled connection between the physical field structure, practical sensing geometry, and theoretical recoverability.
The H/V observation geometry used in our experiments follows the coupled slab-sampling formulation considered in the theory. Importantly, this H/V sampling geometry is motivated by practical ocean sensing, where measurements are collected along horizontal transects and vertical arrays rather than at freely selected spatial locations. The low-rank assumption is also motivated by the structure of acoustic fields. We conducted an additional experiment to examine the low-rank properties of the angular Fourier coefficient fields computed directly from all 80 complete reference fields. The results reveal substantial spectral concentration: spatial rank 5 retains 32.6% of the energy, while ranks 52 and 69 retain 90% and 95%, respectively, for 200 × 200 matrices. These findings provide empirical evidence of approximate low-rank structure, while the additional energy captured at higher ranks reflects the rich spatial variations in realistic acoustic fields. This structural characterization supports the relevance of the theoretical analysis and motivates a natural extension toward more flexible spatial modeling.
Accordingly, TIDAL retains the physics-derived S/c representation and learns dense spatial components that naturally exhibit low-rank structure, as evidenced by Fig. 3(a), while also capturing richer spatial variations without imposing an explicit rank constraint. The resulting architecture preserves the structured spatial–angular coupling and accommodates complex interference patterns, as further supported by the rank-constrained ablations in the original paper. The theoretical guarantee characterizes the exact low-rank specialization under its sufficient conditions, while the practical architecture extends the same representation to more complex acoustic fields. The scope of the theoretical conditions and their connection to the practical architecture are now clarified in Secs. 4.2 and 5, with supporting analyses in Sec. 6.3 (Fig. 3).
The S/c structure originates from the spatial–angular separation in wave propagation physics, providing a physically motivated organization of complex acoustic fields. The physical insight is incorporated through the construction of the representation, while the angular factors remain learnable to capture complex, environment-dependent angular variations.
To further examine this design, we conducted an additional Fixed-Fourier experiment, replacing the learnable angular factors with a prescribed Fourier basis while retaining the S/c decoding structure. The unobserved-region RMSE results are summarized below.
TIDAL consistently achieves better reconstruction accuracy across all three sampling geometries. These results demonstrate the benefit of data-adaptive angular factors within the physics-derived S/c representation, allowing the model to capture complex spatial–angular variations more effectively than the fixed-basis alternative. This comparison is now presented in Sec. 6.5 (Table 2).