pith. sign in

arxiv: 2605.20311 · v1 · pith:NU5R65DXnew · submitted 2026-05-19 · 💻 cs.LG

WaveGraphNet: Physics-Consistent Guided-Wave Damage Localization through Coupled Inverse-Forward Graph Learning

Vinay Sharma , Aditya Bharade , Olga Fink This is my paper

Pith reviewed 2026-05-21 07:17 UTC · model grok-4.3

classification 💻 cs.LG
keywords guided wavedamage localizationgraph learningstructural health monitoringphysics consistentinverse forward couplingcomposite materialssparse sensors
0
0 comments X

The pith

Coupling an inverse damage locator with a forward wave simulator on a transducer graph improves localization accuracy on unseen regions of composite plates.

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

The paper establishes that representing the network of wave transducers as a graph allows a machine learning model to better infer damage locations from limited training examples. By training an inverse branch to guess the damage position and a forward branch to predict how that position would alter the wave energy along each path, the two branches are coupled so that only positions consistent with the actual measurements are favored. A reader would care because real-world monitoring often has sparse sensors and few known damage cases, so this coupling could make reliable predictions possible without needing data from every possible damage spot. The results show better performance than baselines when testing on damage locations not seen during training.

Core claim

The paper claims that a coupled inverse-forward graph learning framework, with transducers as graph nodes and propagation paths as edges, maps spectral descriptors of differential wave responses to damage locations while using the forward branch to predict path-wise energy deviations, thereby regularizing the location estimates to be consistent with wave propagation physics and yielding improved generalization to held-out regions.

What carries the argument

The coupled inverse-forward graph learning on transducer graphs, where the forward branch regularizes the inverse predictions using predicted wave energy redistribution.

If this is right

  • The model provides stronger localization for sparse sensing setups.
  • It shows improved robustness when extrapolating to damage locations outside the training set.
  • Performance exceeds both non-graph and graph-based baselines in the evaluated benchmark.
  • The physics coupling encourages agreement between inferred locations and underlying wave behavior.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This method could be applied to other inverse problems in structural monitoring where partial physics models exist.
  • Reducing reliance on dense labeled damage data might lower the cost of deploying such systems.
  • Testing on structures with different materials or under operational noise would reveal the limits of the regularization approach.

Load-bearing premise

That the forward model's predictions of wave energy changes can effectively guide the location estimates to match the physics of the measured responses better than training the locator alone on the available examples.

What would settle it

A direct comparison showing that the full coupled model does not outperform an inverse-only graph model on localization error for held-out damage positions would indicate the regularizer adds no benefit.

Figures

Figures reproduced from arXiv: 2605.20311 by Aditya Bharade, Olga Fink, Vinay Sharma.

Figure 1
Figure 1. Figure 1: Overview of the proposed WaveGraphNet framework. (a) Guided-wave SHM setup for a CFRP plate with boundary-mounted transducers, where a defect at location p perturbs the propagating wavefield. (b) Signal￾processing pipeline in which the pristine response is subtracted from the measured signal and the resulting differential response is transformed to the frequency do￾main to obtain the path-wise spectral des… view at source ↗
Figure 2
Figure 2. Figure 2: OGW-1 SHM Plate and spatial train–test splits. (a) [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of localization errors across baseline models. [PITH_FULL_IMAGE:figures/full_fig_p034_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative predictions of WaveGraphNet (Inverse-only). Damage localization results for the test set consisting of seen and unseen dam￾age zones (from the spatially held-out region). Blue circles denote true damage locations, red crosses denote predicted damaged locations, green squares denote predicted undamaged locations, the gray marker denotes the no-damage target pud = [−0.001, −0.001], and black t… view at source ↗
Figure 5
Figure 5. Figure 5: Representative predictions of the coupled [PITH_FULL_IMAGE:figures/full_fig_p037_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Localization behavior of the baseline models. [PITH_FULL_IMAGE:figures/full_fig_p038_6.png] view at source ↗
read the original abstract

Guided-wave structural health monitoring enables damage localization in composite plates using sparse networks of bonded piezoelectric transducers. However, inferring the spatial location of defects from pitch-catch measurements remains weakly constrained when only a limited set of damage locations is available for training. As a result, models trained to predict defect locations may perform well on seen cases but generalize poorly to unseen regions of the structure. This paper proposes WaveGraphNet, a coupled inverse--forward graph learning framework for guided-wave damage localization in Carbon Fiber Reinforced Polymer (CFRP) plates. The sensing layout is explicitly modeled as a graph, where transducers are represented as nodes and measured propagation paths define the graph connectivity. An inverse branch maps graph-structured spectral descriptors of differential guided-wave responses to a damage location, while a forward branch predicts the path-wise energy-deviation patterns of measured wave responses associated with a candidate location. During training, the forward branch serves as a physics-consistent regularizer, discouraging location estimates that are numerically plausible but inconsistent with the measured redistribution of wave-response energy. This coupling encourages agreement between inferred damage coordinates and the underlying wave propagation behavior. Within this benchmark, the proposed graph-based formulation provides a strong localization model for sparse guided-wave sensing and demonstrates improved robustness in extrapolation to held-out regions compared to both non-graph and graph baselines. These results highlight the potential of coupled inverse-forward graph learning as an effective strategy for guided-wave localization under limited spatial coverage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes WaveGraphNet, a coupled inverse-forward graph neural network for localizing damage in CFRP plates from sparse guided-wave pitch-catch measurements. Transducers and propagation paths are modeled explicitly as a graph; an inverse branch maps spectral descriptors of differential responses to candidate damage coordinates while a forward branch predicts path-wise energy-deviation patterns. The forward predictions act as a regularizer during training to penalize location estimates inconsistent with observed wave-energy redistribution, with the authors claiming improved extrapolation to held-out spatial regions relative to non-graph and graph baselines.

Significance. If the reported gains in extrapolation hold under rigorous controls, the work would demonstrate a practical way to inject structural consistency into data-limited inverse problems in structural health monitoring. The explicit graph encoding of the sensor layout is a clear methodological strength that aligns naturally with the sparse, path-based nature of guided-wave sensing.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'improved robustness in extrapolation to held-out regions' is asserted without any quantitative metrics, error bars, or statistical tests; the absence of these numbers makes it impossible to assess whether the observed gains exceed baseline variability or are load-bearing for the physics-consistency argument.
  2. [Abstract] Abstract: the forward branch is described as predicting 'path-wise energy-deviation patterns' from candidate locations and serving as a 'physics-consistent regularizer,' yet no indication is given whether this branch is derived from first-principles wave equations or is itself a learned model trained on the same limited damage-location dataset; if the latter, the coupling reduces to multi-task regularization whose generalization benefit must be demonstrated separately from any physics enforcement.
minor comments (1)
  1. [Abstract] The abstract would benefit from a concise statement of the number of transducers, the size of the training damage-location set, and the precise form of the coupling loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'improved robustness in extrapolation to held-out regions' is asserted without any quantitative metrics, error bars, or statistical tests; the absence of these numbers makes it impossible to assess whether the observed gains exceed baseline variability or are load-bearing for the physics-consistency argument.

    Authors: We agree that the abstract would benefit from explicit quantitative support for the extrapolation claim. The full manuscript reports mean localization errors with standard deviations across repeated trials, along with direct comparisons to non-graph and graph baselines for held-out spatial regions. In the revised version we will update the abstract to include representative metrics (e.g., percentage error reduction and associated variability) so that the strength of the reported gains is immediately evident. revision: yes

  2. Referee: [Abstract] Abstract: the forward branch is described as predicting 'path-wise energy-deviation patterns' from candidate locations and serving as a 'physics-consistent regularizer,' yet no indication is given whether this branch is derived from first-principles wave equations or is itself a learned model trained on the same limited damage-location dataset; if the latter, the coupling reduces to multi-task regularization whose generalization benefit must be demonstrated separately from any physics enforcement.

    Authors: The forward branch is a learned graph model trained jointly on the available damage-location dataset; it is not derived from first-principles wave equations. It learns to predict path-wise energy-deviation patterns from candidate locations and graph connectivity, and the coupling uses these predictions to penalize location estimates that are inconsistent with the observed energy redistribution in the training data. This is therefore a data-driven consistency regularizer rather than an explicit physics enforcement. The generalization benefit of the coupling is demonstrated by the improved extrapolation performance relative to the non-coupled baselines in our experiments. We will add a clarifying sentence in the abstract and a short paragraph in the methods section to make the learned nature of the forward branch explicit. revision: yes

Circularity Check

0 steps flagged

No significant circularity; coupled branches are architectural choices with independent empirical validation

full rationale

The provided abstract and description present WaveGraphNet as a coupled inverse-forward graph model where the forward branch is trained to predict energy-deviation patterns and then used as a regularizer during joint optimization. No equations or self-citations are shown that reduce the claimed physics consistency to a fitted parameter or prior result by construction. The forward branch is not described as a first-principles simulator, but the paper does not claim it is; the regularization effect is presented as an empirical outcome of the joint training objective rather than a definitional identity. The central claim of improved extrapolation therefore rests on benchmark comparisons rather than on any load-bearing self-referential step. This is the normal case of a data-driven architecture whose performance must be judged by external validation, not by internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The physics-consistency regularizer is described at a conceptual level without specifying whether it relies on known wave-propagation equations or additional fitted components.

pith-pipeline@v0.9.0 · 5793 in / 1111 out tokens · 34168 ms · 2026-05-21T07:17:25.490180+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 1 internal anchor

  1. [1]

    G¨ uemes, A

    A. G¨ uemes, A. Fernandez-Lopez, A. R. Pozo, J. Sierra-Perez, Structural health monitoring for advanced composite structures: A review, Journal of Composites Science 4 (1) (2020) 13.doi:10.3390/jcs4010013

    work page doi:10.3390/jcs4010013 2020

  2. [2]

    P. Cawley, Structural health monitoring: Closing the gap between re- search and industrial deployment, Structural Health Monitoring 17 (5) (2018) 1225–1244.doi:10.1177/1475921717750047

    work page doi:10.1177/1475921717750047 2018

  3. [3]

    Ricci, E

    F. Ricci, E. Monaco, N. D. Boffa, L. Maio, V. Memmolo, Guided waves for structural health monitoring in composites: A review and imple- mentation strategies, Progress in Aerospace Sciences 129 (2022) 100790. doi:10.1016/j.paerosci.2021.100790

    work page doi:10.1016/j.paerosci.2021.100790 2022

  4. [4]

    Capineri, A

    L. Capineri, A. Bulletti, Ultrasonic guided-waves sensors and integrated structural health monitoring systems for impact detection and localiza- tion: A review, Sensors 21 (9) (2021) 2929.doi:10.3390/s21092929

    work page doi:10.3390/s21092929 2021

  5. [5]

    Philibert, K

    M. Philibert, K. Yao, M. Gresil, C. Soutis, Lamb waves-based technolo- gies for structural health monitoring of composite structures for air- craft applications, European Journal of Materials 2 (1) (2022) 436–474. doi:10.1080/26889277.2022.2094839

    work page doi:10.1080/26889277.2022.2094839 2022

  6. [6]

    P. Cawley, Guided waves in long range nondestructive testing and structural health monitoring: Principles, history of applications and prospects, NDT & E International 142 (2024) 103026.doi:10.1016/j. ndteint.2023.103026

    work page doi:10.1016/j 2024

  7. [7]

    Tanveer, M

    M. Tanveer, M. U. Elahi, J. Jung, M. M. Azad, S. Khalid, H. S. Kim, Recent advancements in guided ultrasonic waves for structural health monitoring of composite structures, Applied Sciences 14 (23) (2024) 11091.doi:10.3390/app142311091

    work page doi:10.3390/app142311091 2024

  8. [8]

    Sattarifar, T

    A. Sattarifar, T. Nestorovi´ c, Emergence of machine learning techniques in ultrasonic guided wave-based structural health monitoring, Interna- tional Journal of Prognostics and Health Management 13 (1) (2022). doi:10.36001/ijphm.2022.v13i1.3107

    work page doi:10.36001/ijphm.2022.v13i1.3107 2022

  9. [9]

    Y. Liao, X. Qing, Y. Wang, F. Zhang, Damage localization for composite structure using guided wave signals with gramian angular field image 40 coding and convolutional neural networks, Composite Structures 312 (2023) 116871.doi:10.1016/j.compstruct.2023.116871

    work page doi:10.1016/j.compstruct.2023.116871 2023

  10. [10]

    Y. Gao, L. Sun, R. Song, C. Peng, X. Wu, J. Wei, M. Jiang, Q. Sui, L. Zhang, Damage localization in composite structures based on lamb wave and modular artificial neural network, Sensors and Actuators A: Physical 370 (2024) 115644.doi:10.1016/j.sna.2024.115644

    work page doi:10.1016/j.sna.2024.115644 2024

  11. [11]

    M. M. Azad, O. Munyaneza, J. Jung, J. W. Sohn, J.-W. Han, H. S. Kim, Damage localization and severity assessment in composite structures using deep learning based on lamb waves, Sensors 24 (24) (2024) 8057. doi:10.3390/s24248057

    work page doi:10.3390/s24248057 2024

  12. [12]

    Lomazzi, M

    L. Lomazzi, M. Giglio, F. Cadini, Towards a deep learning-based uni- fied approach for structural damage detection, localisation and quan- tification, Engineering Applications of Artificial Intelligence 121 (2023) 106003.doi:10.1016/j.engappai.2023.106003

    work page doi:10.1016/j.engappai.2023.106003 2023

  13. [13]

    The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009.doi:10.1109/TNN.2008.2005605

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, G. Monfardini, The graph neural network model, IEEE Transactions on Neural Networks 20 (1) (2009) 61–80.doi:10.1109/TNN.2008.2005605

    work page doi:10.1109/tnn.2008.2005605 2009

  14. [14]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Neural message passing for quantum chemistry, Proceedings of the 34th Inter- national Conference on Machine Learning 70 (2017) 1263–1272

    work page 2017

  15. [15]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph con- volutional networks, in: International Conference on Learning Repre- sentations, 2017. URLhttps://openreview.net/forum?id=SJU4ayYgl

    work page 2017

  16. [16]

    Veli ˇkovi´ c, G

    P. Veli ˇkovi´ c, G. Cucurull, A. Casanova, A. Romero, P. Li` o, Y. Bengio, Graph attention networks, in: International Conference on Learning Representations, 2018. URLhttps://openreview.net/forum?id=rJXMpikCZ

    work page 2018

  17. [17]

    Sharma, O

    V. Sharma, O. Fink, A physics-informed graph neural network con- serving linear and angular momentum for dynamical systems, Nature Communications (2026). 41

    work page 2026

  18. [18]

    Bloemheuvel, J

    S. Bloemheuvel, J. van den Hoogen, M. Atzmueller, A computa- tional framework for modeling complex sensor network data using graph signal processing and graph neural networks in structural health monitoring, Applied Network Science 6 (1) (2021).doi:10.1007/ s41109-021-00438-8

    work page 2021

  19. [19]

    M. A. Cheema, M. Z. Sarwar, V. C. Gogineni, D. Cantero, P. S. Rossi, Computationally efficient structural health monitoring using graph sig- nal processing, IEEE Sensors Journal 24 (7) (2024) 11895–11905.doi: 10.1109/JSEN.2024.3366346

    work page doi:10.1109/jsen.2024.3366346 2024

  20. [20]

    S. Wang, Z. Luo, P. Shen, H. Zhang, Z. Ni, Graph-in-graph convolu- tional network for ultrasonic guided wave-based damage detection and localization, IEEE Transactions on Instrumentation and Measurement 71 (2022) 1–13.doi:10.1109/TIM.2022.3144732

    work page doi:10.1109/tim.2022.3144732 2022

  21. [21]

    L. Sun, R. Song, J. Wei, Y. Gao, C. Peng, L. Fan, M. Jiang, L. Zhang, Physics-augmented spatial-temporal graph convolutional network for damage localization using ultrasonic guided waves, Mechanical Systems and Signal Processing 221 (2024) 111738.doi:10.1016/j.ymssp.2024. 111738

    work page doi:10.1016/j.ymssp.2024 2024

  22. [22]

    Rautela, S

    M. Rautela, S. Gopalakrishnan, Ultrasonic guided wave based structural damage detection and localization using model assisted convolutional and recurrent neural networks, Expert Systems with Applications 167 (2021) 114189.doi:10.1016/j.eswa.2020.114189

    work page doi:10.1016/j.eswa.2020.114189 2021

  23. [23]

    Y. Song, S. Shan, Y. Zhang, L. Cheng, Physics-guided neural network for structural health monitoring with lamb waves through boundary reflection elimination, Structural Health Monitoring 25 (2) (2026) 1219– 1236.doi:10.1177/14759217241305050

    work page doi:10.1177/14759217241305050 2026

  24. [24]

    Ongie, A

    G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, R. Willett, Deep learning techniques for inverse problems in imaging, IEEE Journal on Selected Areas in Information Theory 1 (1) (2020) 39–56.doi:10.1109/JSAIT.2020.2991563

    work page doi:10.1109/jsait.2020.2991563 2020

  25. [25]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of 42 Computational Physics 378 (2019) 686–707.doi:10.1016/j.jcp.2018. 10.045

    work page doi:10.1016/j.jcp.2018 2019

  26. [26]

    Del Priore, L

    E. Del Priore, L. Lampani, Real-time damage detection and localization on aerospace structures using graph neural networks, Journal of Sensor and Actuator Networks 14 (5) (2025) 89.doi:10.3390/jsan14050089

    work page doi:10.3390/jsan14050089 2025

  27. [27]

    K. F. Niresi, Z. Zhang, O. Fink, Rins-t: Robust implicit neural solvers for time-series linear inverse problems, IEEE Transactions on Instru- mentation and Measurement 74 (2025) 1–14

    work page 2025

  28. [28]

    O. Fink, V. Sharma, I. Nejjar, L. Von Krannichfeldt, S. Garmaev, Z. Zhang, A. Wei, G. Frusque, F. Forest, M. Zhao, et al., From physics to machine learning and back: Part i-learning with inductive biases in prognostics and health management, Reliability Engineering & System Safety (2026) 112213

    work page 2026

  29. [29]

    R. Song, L. Sun, Y. Gao, J. Wei, C. Peng, L. Fan, M. Jiang, Unsuper- vised temperature-compensated damage localization method based on damage to baseline autoencoder and delay-based probabilistic imaging, Mechanical Systems and Signal Processing 230 (2025) 112649

    work page 2025

  30. [30]

    J. E. Michaels, Detection, localization and characterization of damage in plates with an in situ array of spatially distributed ultrasonic sensors, Smart Materials and Structures 17 (3) (2008) 035035.doi:10.1088/ 0964-1726/17/3/035035

    work page 2008

  31. [31]

    Z. Su, L. Ye, Y. Lu, Guided lamb waves for identification of damage in composite structures: A review, Journal of Sound and Vibration 295 (3–

    work page

  32. [32]

    (2006) 753–780.doi:10.1016/j.jsv.2006.01.020

    work page doi:10.1016/j.jsv.2006.01.020 2006

  33. [33]

    J. Moll, J. Kathol, C.-P. Fritzen, M. Moix-Bonet, M. Rennoch, M. Ko- erdt, A. S. Herrmann, M. G. R. Sause, M. Bach, Open guided waves: online platform for ultrasonic guided wave measurements, Struc- tural Health Monitoring 18 (5–6) (2019) 1903–1914.doi:10.1177/ 1475921718817169

    work page 2019

  34. [34]

    Z. Wang, W. Yan, T. Oates, Time series classification from scratch with deep neural networks: A strong baseline, in: 2017 International joint conference on neural networks (IJCNN), IEEE, 2017, pp. 1578–1585. 43

    work page 2017

  35. [35]

    Hochreiter, Long short-term memory, Neural Computation MIT- Press (1997)

    S. Hochreiter, Long short-term memory, Neural Computation MIT- Press (1997)

    work page 1997

  36. [36]

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zam- baldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, C. Gulcehre, F. Song, A. Ballard, J. Gilmer, G. E. Dahl, A. Vaswani, K. R. Allen, C. Nash, V. Langston, C. Dyer, N. Heess, D. Wierstra, P. Kohli, M. Botvinick, O. Vinyals, Y. Li, R. Pascanu, Relational inductive bi...

    work page internal anchor Pith review Pith/arXiv arXiv 2018