Pith. sign in

REVIEW 4 major objections 9 minor 25 references

PyAWD: A Library for Generating Large Synthetic Datasets of Acoustic Wave Propagation

T0 review · 4 major / 9 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read PyAWD is a Python library for generating large, labeled acoustic-wave datasets for machine learning, with an epicenter-retrieval demonstration from two sensors.

desk verdict PyAWD is a useful, genuinely PyTorch-integrated synthetic wave simulator with real code, but the paper overstates its novelty, validates only in a closed loop, and prints a wave equation with incorrect units that needs checking. read the letter →

arxiv 2411.12636 v2 pith:GIELHQJT submitted 2024-11-19 cs.LG

Pascal Tribel , Gianluca Bontempi This is my paper · ORCID
classification cs.LG
keywords wavesimulationsyntheticseismicdataacousticpropagationpytorchdatasetsmachinelearningepicenterretrievalbudgetingheterogeneousmedia
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PyAWD is a proposed answer to a concrete gap: seismic data is sparse and expensive to collect, yet no existing tool produces large, labeled datasets of simulated ground motion for machine learning. The paper claims PyAWD fills that gap by solving acoustic wave propagation in heterogeneous media and exposing each simulation as a tensor-based dataset that deep-learning pipelines can consume directly, with full control over wave speed, attenuation, external forces, discretization, and medium composition. To show the data are not trivial, the authors run an epicenter-retrieval study in which models trained on raw two-sensor seismograms locate the source in a heterogeneous field better than simple baselines, and they use the same generator to study data budgeting—how much data, how many sensors, and how much noise the task tolerates. A sympathetic reader would take the paper's contribution to be an enabling tool plus evidence that synthetic wave data supports nontrivial seismic ML tasks.

What carries the argument

The load-bearing object is the anisotropic nondispersive acoustic wave equation $$\frac{$d^{2}$u}{$dt^{2}$} = c\,\$nabla^{2}$ u - \$\alpha$\,\frac{du}{dt} + f$$, where $u$ is the displacement field, $c=c(x,y,t)$ is a spatio-temporally varying propagation-speed field, $\alpha$ an attenuation factor, and $f$ an external force. PyAWD discretizes this equation with finite differences and wraps the solver so a user can generate many simulations with different media, sources, and sensor placements; each simulation is returned as a sequence of seismograms recorded by 'interrogators,' the library's abstraction for seismometers. The demonstration's physical bet is that wave self-interference in a heterogeneous medium encodes information about source location, so raw multi-channel seismograms can be fed directly to a regression model.

What would settle it

Train a model on PyAWD acoustic data and test it on real single-station seismograms from a region with a known velocity model; if the model's normalized mean squared error is no better than predicting the training-set mean, the central claim that these synthetic data support useful seismic machine learning is falsified.

Watch

Extended reading notes

Core claim

The central claim is that PyAWD is the first library that couples customizable finite-difference simulation of acoustic wave propagation with on-demand generation of ML-ready datasets, and that this is enough to support realistic seismic inverse problems. In the demonstration, the library generates 4096 simulations in a 256×256 discretized Marmousi velocity model, each running 10 seconds with two interrogators recording horizontal and vertical motion at 100 hertz; randomized epicenters make the training set cover only 5.57 percent of possible grid positions. Temporal convolutional networks and Extra Trees both retrieve epicenters with normalized mean squared error well below the constant baseline, and the generated data reveal spatial error patterns and data-budgeting curves. The paper's own caveat is that the wave is acoustic rather than elastic, so the usual P/S arrival-time distance cue is absent from these signals.

Load-bearing premise

The load-bearing premise is that acoustic wave propagation through a prescribed velocity model is an adequate proxy for real seismic ground motion, so that performance measured on simulated data transfers to real seismograms.

Editorial extensions

If this is right

  • Any researcher can create arbitrarily large, labeled sets of wave-propagation simulations without deploying physical seismometers, making data-scarce seismic ML problems tractable as a first pass.
  • Because the output is a standard tensor-based dataset, synthetic simulations can be streamed on the fly, augmented, shuffled, and batched through existing deep-learning data pipelines.
  • In the epicenter demo, two interrogators suffice for a temporal convolutional network to beat simple baselines, suggesting that one- or two-station arrays can carry more location information than classical triangulation assumes.
  • The same generator yields learning curves that quantify how many samples, how many sensors, and how much noise the task tolerates—inputs to a data-budgeting decision before field deployment.
  • Error heatmaps over all possible source coordinates expose where a model is weakest (near field boundaries), which can guide sensor placement and model improvement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural test the paper does not run is to swap the acoustic solver for an elastic one; because elastic waves carry P/S arrival-time differences, the two-interrogator advantage should reappear or strengthen, and its absence in the acoustic demo could be checked directly.
  • The data-budgeting curves would become actionable only after a real-data validation step: compare a model trained on PyAWD data and finetuned on a small real set against one trained on real data alone, which the paper leaves as future work.
  • The tool's design implies a benchmark suite: fixed velocity models, source distributions, and interrogator layouts could standardize comparisons of epicenter-location architectures, an extension not proposed in the paper.
  • Because the simulator controls the medium exactly, it supports controlled experiments on sensor placement and noise robustness that real deployments cannot provide, and the results of those experiments could inform physical deployment choices.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. This paper presents PyAWD, an open-source Python library that generates synthetic acoustic-wave-propagation datasets for machine learning. The library uses Devito to solve scalar and vectorial acoustic wave equations on two- and three-dimensional heterogeneous media, wraps the outputs as PyTorch datasets, and provides presets (e.g., the Marmousi field) and an interrogator abstraction for emulating seismometers. The authors demonstrate the library on a 2D epicenter-retrieval task with two interrogators, comparing seven ML regressors, and additionally study data budgeting (training-set size, number of interrogators, and input noise). The paper argues that PyAWD fills a gap by producing large, PyTorch-compatible synthetic seismic datasets that can support ML research in seismology.

Significance. If the implementation is correct, PyAWD addresses a real practical gap: few integrated tools generate machine-learning-scale, PyTorch-ready synthetic seismic-like datasets with flexible control over sources, media, and interrogation points. The library is open-source, documented with notebooks, and the epicenter-retrieval demo shows a nontrivial regression problem in which raw 1000-step two-channel seismograms yield NMSE well below the constant baseline. The data-budgeting experiments illustrate a useful application. However, the paper's core validation is closed-loop, and the printed wave equation contains a unit/convention error that must be resolved before the physical-fidelity claims can be accepted. The manuscript is therefore promising but needs substantial revision.

major comments (4)
  1. [§2, Eq. (1)] The stated units for c are incorrect. The text says 'c describes the wave propagation speed field (a scalar field, in s²/m²)'. For Eq. (1), d²u/dt² = c∇²u, with u in meters and t in seconds, the coefficient c must have units m²/s² if it multiplies ∇²u; if c is intended to be a speed in m/s, the equation should be c²∇²u. The closed-loop experiment in Section 3 cannot detect a wrong but self-consistent convention, because a uniformly scaled speed still produces learnable arrival-time patterns. Please verify that the Devito operator matches the printed formula, correct the equation and/or units, and state explicitly whether the library's 'wave speed' parameter is phase speed or its square.
  2. [§3.1–§3.3] The validation is entirely closed-loop: the same simulator generates both training and test data, and the test epicenters are drawn from the same distribution as the training epicenters. The statement in §3.1 that the experiment 'assess[es] that PyAWD does not oversimplify reality' is therefore not supported. The reported NMSE measures internal consistency under the simulator's forward model, not agreement with real seismograms or with an independent reference. Please add an external sanity check, for example comparing simulated arrival times with analytic travel times in a homogeneous medium, or comparing against an independent finite-difference/finite-element solver for a heterogeneous model, and report the discrepancy.
  3. [§3.2, Figure 6] The main accuracy comparison reports that the training/testing procedure was run 10 times, but Figure 6 shows no error bars, confidence intervals, or statistical tests. Without variability information, claims such as 'the Temporal Convolutional Neural Network and the Extra Trees largely outperform the other models' (§4) are not substantiated. Please add error bars or per-repetition scatter points, and ideally a paired significance test across the 10 splits.
  4. [§2, Eq. (2)] The finite-difference discretization is presented without any discussion of numerical stability (CFL condition), grid dispersion, or boundary conditions. For a dataset-generation library whose outputs are used as ground truth, the absence of a convergence/stability check leaves open the possibility that generated samples contain numerical artifacts. Please report the scheme's order of accuracy, the CFL criterion used by the implementation, and a convergence test on a case with an analytic solution.
minor comments (9)
  1. [§2, Eq. (1)] Equation (1) uses ordinary derivatives d²u/dt² for what is a partial differential equation; use ∂²u/∂t² and ∂²u/∂x² throughout.
  2. [§2, text after Eq. (1)] The equation is labeled 'anisotropic' but the coefficient c is a scalar field. Please clarify what anisotropy means in this context, or use the conventional terminology.
  3. [§2, Eq. (2)] The staggered-grid offsets hx and hy in Eq. (2) are not defined, and the finite-difference stencil is not fully specified. Please provide the explicit stencil or a reference.
  4. [§3.1] The boundary conditions of the simulations are never stated. Since the epicenter-retrieval motivation relies on wave reflections from heterogeneities, specifying whether boundaries are absorbing or reflecting is essential for reproducibility.
  5. [§3.2, Figure 6] Figure 6 appears to use a logarithmic vertical axis, but the axis label is not explicit. State the scale explicitly once error bars are added.
  6. [Abstract] The claim that 'no tool allows the generation of large datasets containing simulated measurements of the ground motion' is too strong given the cited Synthoseis and full-waveform-inversion works. Please qualify the novelty, e.g., 'no open-source tool that combines 2D/3D heterogeneous acoustic simulation with PyTorch dataset integration and automatic initial-condition randomization.'
  7. [§3.1–§3.2] The dataset sizes are inconsistent: §3.1 says 3584 training and 512 test samples, while §3.2 mentions a training set of 3284 samples and a random split into 80%/20% subsets. Please reconcile these numbers and describe the exact split procedure.
  8. [§3.3, Figure 9a] The statement '62 samples seem to be enough to beat the constant average model' is presented without a statistical test or a methodology for choosing this threshold. Please indicate how this number is derived.
  9. [Throughout] There are multiple typos and inconsistencies in naming: 'Pytorch' should be 'PyTorch', 'ExtraTrees' vs 'Extra Trees' should be unified, and in the code-availability section 'PyA WD' should be 'PyAWD'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: PyAWD is a wrapper around an external finite-difference solver; the ML demo is a closed-loop internal-consistency check, not a prediction derived from fitted inputs.

full rationale

The paper contains no load-bearing circular step. The physical solver is Devito, an external finite-difference package, and the acoustic wave equation is cited to Demanet (2015); neither is a self-citation. Epicenter coordinates, wave speed fields, source amplitudes, and delays are drawn from random distributions and used as labels in a conventional supervised train/test split; no parameter is fitted to the target quantity and then renamed a prediction. The epicenter-retrieval experiment is closed-loop in the sense that PyAWD generates both training and test data, so the measured NMSE demonstrates learnability and internal consistency of the synthetic datasets rather than transfer to real seismograms; the paper explicitly acknowledges this in Section 4 ('One primary limitation is the reliance on synthetic datasets...'), which is a generalizability caveat, not a circular reduction. The dimensional convention in Eq. (1) (c labeled in s^2/m^2 for a wave equation) is a physical-correctness concern, not a circularity concern. No self-citation chain is load-bearing, and no uniqueness theorem or ansatz is smuggled in via the authors' prior work.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the fidelity of the acoustic wave simulation and the representativeness of the Marmousi demo. No new physical entities are introduced, and no parameters are fitted to external data, but several experimental settings are hand-chosen and shape the reported results.

free parameters (3)
  • Number of interrogators = 2
    Hand-chosen to demonstrate epicenter retrieval with fewer than three stations; the data budgeting results in Figure 9b depend on this choice.
  • Training set size = 3584
    Hand-chosen so epicenter coverage is 5.57% of the 256x256 grid; the '62 samples enough' result in Section 4 depends on the training-set construction.
  • Epicenter sampling and force ranges = uniform over field; amplitude [0.5,1.5); delay [0,5)s
    Hand-chosen experimental design parameters that shape the dataset but are not fitted to an external target.
assumptions (3)
  • domain assumption The scalar and vectorial acoustic wave equation with attenuation is an adequate model of seismic ground motion for ML training.
    The entire library simulates acoustic waves; the paper acknowledges real earthquakes involve elastic P and S waves with different speeds (Section 3).
  • standard math Devito's finite difference solver is stable and accurate for the parameter ranges used.
    No stability analysis or verification against analytic solutions is provided; the paper relies on Devito defaults (Methods, Section 2).
  • domain assumption The Marmousi velocity model is a representative heterogeneous medium for seismic ML studies.
    Used as the preset in the demo; its realism is assumed, not verified independently (Figure 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PyAWD: A Library for Generating Large Synthetic Datasets of Acoustic Wave Propagation." pith.science (2026). https://pith.science/paper/GIELHQJT

@misc{pith2026241112636,
  author       = {Pith},
  title        = {Pith review of: PyAWD: A Library for Generating Large Synthetic Datasets of Acoustic Wave Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GIELHQJT}},
  note         = {Machine review of arXiv:2411.12636}
}
read the original abstract

Seismic data is often sparse and unevenly distributed due to the high costs and logistical challenges associated with deploying physical seismometers, limiting the application of Machine Learning (ML) in earthquake analysis. While simulation methods exist, no tool allows the generation of large datasets containing simulated measurements of the ground motion. To address this gap, we introduce PyAWD, a Python library designed to generate high-resolution synthetic datasets simulating spatio-temporal acoustic wave propagation in both two-dimensional and three-dimensional heterogeneous media. By allowing fine control over parameters such as the wave speed, external forces, spatial and temporal discretization, and media composition, PyAWD enables the creation of ML-scale datasets that capture the complexity of seismic wave behavior. We illustrate the library's potential with an epicenter retrieval task, showcasing its suitability for designing complex, accurate seismic problems that require advanced ML approaches in the absence or lack of dense real-world data. We also show the usefulness of our tool to tackle the problem of data budgeting in the framework of epicenter retrieval.

Figures

Figures reproduced from arXiv: 2411.12636 by the authors.

Figure 1
Figure 1. Example of scalar wave propagation simulation generated by [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. The Marmousi field, from (Versteeg (1994)). The darker the color, the lower the propagation [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Histogram of both x and y coordinates of the epicenters in the training set [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Example of propagating wave in the Marmousi field, with two interrogators. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Example of interrogators response, which follows the simulation shown in figure 4 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Normalized Mean Squared Error of the predicted coordinates for each model. The training/testing [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Comparison of predicted versus expected coordinates for the testing set, using the Temporal [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Heatmap of the distance between the expected and predicted coordinates for every possible [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Data budgeting and noise sensitivity study. The predictions are averaged over 5 different instances [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 9 canonical work pages

  1. [2]

    Noda, S., Yamamoto, S., Sato, S., Iwata, N., Korenaga, M., Ashiya, K.,

    URL: https://seismica.library.mcgill.ca/ article/view/368, doi:10.26443/seismica.v2i1.368. Noda, S., Yamamoto, S., Sato, S., Iwata, N., Korenaga, M., Ashiya, K.,

    work page doi:10.26443/seismica.v2i1.368

  2. [3]

    Mohr, F., van Rijn, J.N.,

    doi: 10.7717/peerj-cs.103. Mohr, F., van Rijn, J.N.,

    doi:10.7717/peerj-cs.103

  3. [4]

    Louboutin, M., Lange, M., Luporini, F., Kukreja, N., Witte, P.A., Herrmann, F.J., Velesko, P., Gorman, G.J.,

    Synthetic ground motions in heterogeneous geologies: the hemew-3d dataset for scientific machine learning doi: 10.5194/essd-2023-470. Louboutin, M., Lange, M., Luporini, F., Kukreja, N., Witte, P.A., Herrmann, F.J., Velesko, P., Gorman, G.J.,

    work page doi:10.5194/essd-2023-470 2023

  4. [6]

    18 Singh, S.P., Silwal, V.,

    doi: 10.1186/s43074-025-00160-z . 18 Singh, S.P., Silwal, V.,

    work page doi:10.1186/s43074-025-00160-z

  5. [9]

    CoRR abs/2201.12150

    Learning curves for decision making in supervised machine learning - A survey. CoRR abs/2201.12150. URL: https://arxiv.org/abs/2201.12150, arXiv:2201.12150. Mousavi, S., Weiqiang, Z., Sheng, Y., Beroza, G., 2019a. Cred: A deep residual network of convolutional and recurrent units for earthquake signal detection. Scientific Reports

    arXiv

  6. [10]

    Mousavi, S.M., Beroza, G.C.,

    doi: 10.1038/s41598-019-45748-1 . Mousavi, S.M., Beroza, G.C.,

    doi:10.1038/s41598-019-45748-1

  7. [12]

    Luporini, F., Louboutin, M., Lange, M., Kukreja, N., Witte, P., H¨ uckelheim, J., Yount, C., Kelly, P.H.J., Herrmann, F.J., Gorman, G.J.,

    doi: 10.5194/gmd-12-1165-2019 . Luporini, F., Louboutin, M., Lange, M., Kukreja, N., Witte, P., H¨ uckelheim, J., Yount, C., Kelly, P.H.J., Herrmann, F.J., Gorman, G.J.,

    doi:10.5194/gmd-12-1165-2019 2019

  8. [15]

    Demanet, L.,

    doi: 10.3389/fpsyg.2024.1408073. Demanet, L.,

    arXiv 2024

Show all 25 references
  1. [16]

    Richardson, A.,

    doi: 10.1126/sciadv.1700578. Richardson, A.,

    doi

  2. [17]

    Earth Science Informatics

    Machine learning for earthquake prediction: a review (2017–2021). Earth Science Informatics

    2017

  3. [18]

    Ross, Z., Meier, M.A., Hauksson, E.,

    doi: 10.1007/s12145-023-00991-z . Ross, Z., Meier, M.A., Hauksson, E.,

    doi

  4. [20]

    IEEE Geoscience and Remote Sensing Letters 1, 1–1

    Eqconvmixer: A deep learning approach for earthquake location from single-station waveforms. IEEE Geoscience and Remote Sensing Letters 1, 1–1. doi: 10.1109/LGRS.2023.3312324. Shao, L.Y., Zhang, J., Chen, X., Xu, D., Gu, H., Mu, Q., Yu, F., Liu, S., Shi, X., Sun, J., Huang, Z....

    2023

  5. [22]

    2023.10.002

    doi: 10.1016/j.aiig. 2023.10.002. T¨ urkmen, M., Meral, S., Yilmaz, B., Cikis, M., Akag¨ und¨ uz, E., Tileylioglu, S.,

    2023 doi

  6. [24]

    Seismological Research Let- ters 93, 1695–1709

    SeisBench—A Toolbox for Machine Learning in Seismology. Seismological Research Let- ters 93, 1695–1709. URL: https://doi.org/10.1785/0220210324, doi: 10.1785/0220210324, arXiv:https://pubs.geoscienceworld.org/ssa/srl/article-pdf/93/3/1695/5596371/srl-2021324.1.pdf. Zhao, X., L...

    doi

  7. [25]

    Zhu, W., Mousavi, S.M., Beroza, G.C.,

    Data Budgeting for Machine Learning URL: http://arxiv.org/abs/ 2210.00987, doi:10.48550/arXiv.2210.00987. Zhu, W., Mousavi, S.M., Beroza, G.C.,

    doi arXiv

  8. [46]

    maintainers, T., contributors,

    doi: 10.1145/3374916. maintainers, T., contributors,

    doi

  9. [57]

    doi: 10.1109/TGRS.2019.2926772. 19

    2019

  10. [123]

    Saad, H., M

    doi: 10.1029/2017JB015251. Saad, H., M. Saad, O., Soliman, M., Chen, Y., Youness, H.,

    doi

  11. [218]

    Geurts, P., Ernst, D., Wehenkel, L.,

    doi: 10.1093/gji/ ggz226. Geurts, P., Ernst, D., Wehenkel, L.,

    doi

  12. [1994]

    The Leading Edge 13, 927–936

    The marmousi experience: Velocity model determination on a synthetic complex data set. The Leading Edge 13, 927–936. doi: 10.1190/1.1437051. Woollam, J., M¨ unchmeyer, J., Tilmann, F., Rietbrock, A., Lange, D., Bornstein, T., Diehl, T., Giunchi, C., Haslinger, F., Jozinovi´ c,...

    doi

  13. [2012]

    Earth Planets and Space 64, 305–308

    Improvement of back- azimuth estimation in real-time by using a single station record. Earth Planets and Space 64, 305–308. doi:10.5047/eps.2011.10.005. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.,

    2011 doi

  14. [2018]

    URL: https://arxiv.org/abs/1801.07232, arXiv:1801.07232

    Seismic full-waveform inversion using deep learning tools and techniques. URL: https://arxiv.org/abs/1801.07232, arXiv:1801.07232. Ridzwan, N., Md Yusoff, S.H.,

    arXiv

  15. [2022]

    Interpretation 10, SE31– SE39

    Synthetic seismic data for training deep learning networks. Interpretation 10, SE31– SE39. doi: 10.1190/INT-2021-0193.1 . Meurer, A., Smith, C.P., Paprocki, M., ˇCert ´ ık, O., Kirpichev, S.B., Rocklin, M., Kumar, A., Ivanov, S., Moore, J.K., Singh, S., Rathnayake, T., Vig, S....

    2021 doi

  16. [2023]

    Annual Review of Earth and Planetary Sciences 51, 105–129

    Machine learning in earthquake seismology. Annual Review of Earth and Planetary Sciences 51, 105–129. URL: https://www.annualreviews.org/content/journals/10.1146/ annurev-earth-071822-100323 , doi:https://doi.org/10.1146/annurev-earth-071822-100323 . Mousavi, S.M., Sheng, Y., ...

    doi

  17. [2024]

    URL: https://arxiv.org/abs/2403.18116, arXiv:2403.18116

    Quakeset: A dataset and low-resource models to monitor earthquakes through sentinel-1. URL: https://arxiv.org/abs/2403.18116, arXiv:2403.18116. Carlini, A., Bordeau, C., Ambard, M.,

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.