Pith. sign in

REVIEW 3 major objections 6 minor 24 references

A physics-informed neural network trained with a three-stage loss schedule reproduces coupled electro-elastic wave propagation to within 2.3% (displacement) and 4.9% (electric potential) error on a 1D piezoelectric benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A PINN is applied to a 1D piezoelectric wave system and reports 2.3%/4.9% L2 errors, but the 'exact solution' used for validation is inconsistent with the governing equations.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection This is a clearly-written PINN application to a 1D piezoelectric wave problem, but the validation benchmark falls apart: the advertised exact solution does not satisfy the residual equations as written. the 3 major comments →

arxiv 2602.13811 v2 pith:JEECY46B submitted 2026-02-14 cs.NE cs.LGphysics.comp-ph

A Unified Physics-Informed Neural Network for Modeling Coupled Electro- and Elastodynamic Wave Propagation Using Three-Stage Loss Optimization

Suhas Suresh Bharadwaj , Reuben Thomas Thovelil This is my paper
classification cs.NE cs.LGphysics.comp-ph MSC 68T0774F1565M99
keywords physics-informed neural networkspiezoelectricityelectro-elastodynamicspartial differential equationsmesh-free solvershard constraintsthree-stage loss optimizationstanding wave benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper attempts to show that a physics-informed neural network, trained with no labeled interior data and only the governing PDE residuals as soft constraints, can solve a coupled electro- and elastodynamic system—specifically one-dimensional linear piezoelectricity in stress-charge form. The authors report global relative L2 errors of 2.34% for mechanical displacement and 4.87% for electric potential against a known standing-wave solution, and argue this validates PINNs as effective mesh-free solvers for coupled time-dependent multiphysics problems. The central demonstration is that a careful combination of hard boundary/initial constraints and a three-stage optimizer (Adam, then AdamW, then L-BFGS) lets a standard feedforward network learn both fields simultaneously. The work matters because coupled multiphysics systems are a known pain point for PINNs, and this provides a reproducible benchmark with a concrete training recipe.

Core claim

On the paper's own terms, a fully-connected network mapping space-time coordinates (x, t) to mechanical displacement u and electric potential φ can learn the coupled response of a 1D piezoelectric bar. Three design choices carry the result: hard constraints embed the Dirichlet boundary conditions and initial conditions through output transformations u_constrained = x(1−x)u_raw + sin(πx)(1−t) (and analogously for φ), removing boundary-condition error from the loss; the physics is enforced by two PDE residuals (elastodynamic r1 and electrodynamic r2) evaluated at 20,000 collocation points; and a three-stage optimizer—Adam for rapid initial descent, AdamW with weight decay for regularization, L

What carries the argument

The two PDE residuals—r1 = ρ u_tt − (cE u_xx − e33 φ_xx) and r2 = ε0 φ_tt + e33 u_xx + ε0 φ_xx—are the physics-informed constraints that define the loss. The hard-constraint output transformation, which multiplies the raw network output by x(1−x) and adds the initial-profile term sin(πx)(1−t), algebraically forces boundary and initial conditions, so the network's capacity is spent entirely on minimizing interior residuals rather than on penalty-based constraint satisfaction. The three-stage optimizer (Adam, AdamW, L-BFGS) is the procedural mechanism that makes this loss minimization converge in practice.

Load-bearing premise

The load-bearing assumption is that the closed-form solutions u = sin(πx)cos(πt) and φ = 0.5 sin(πx)cos(πt) exactly solve the residual equations used in training; if they do not, the reported L2 errors tell us nothing about how well the network solves the PDE system.

What would settle it

Evaluate the residuals r1 and r2 at the claimed exact solution (11)–(12) using the material parameters the code actually uses (which the paper does not state). If either residual is nonzero for those parameters, the benchmark solution is not a solution of the trained system, and the reported 2.34% and 4.87% errors would be measuring approximation of a different problem.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported accuracy holds, PINNs can serve as mesh-free surrogates or fast pre-screeners for 1D coupled piezoelectric wave problems, with errors in the few-percent range over the full space-time domain.
  • The three-stage optimizer order—rapid Adam descent, AdamW regularization, then L-BFGS polishing—offers a reusable recipe for training PINNs on coupled PDE systems where single-optimizer training stalls.
  • The hard-constraint approach guarantees boundary errors below 1e−6, demonstrating that boundary conditions can be enforced exactly rather than learned as soft penalties.
  • The observed error growth over time and the coupling-induced amplification from displacement to electric potential are concrete limitations that any future PINN architecture for coupled multiphysics must address.
  • These results provide a quantitative benchmark (2.34% and 4.87% relative L2 errors on a 450×450 grid) against which more advanced PINN variants can be compared.

Where Pith is reading between the lines

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

  • My inference: the benchmark's validity hinges on the claimed closed-form solution actually satisfying the residual equations with the material parameters used in the code; since the paper omits those parameters and the residual r2 appears to have a typo (ε0 vs εS from the constitutive equation), the reported errors may be measuring approximation of a different PDE system than the one stated.
  • My inference: the error amplification mechanism described here predicts that in higher-dimensional or stronger-coupling regimes, the electric-field error will grow disproportionately; a natural extension is to apply the same three-stage recipe to a 2D piezoelectric geometry and watch whether the electric potential error exceeds 10%.
  • My inference: the observed saturation of errors at late times suggests a residual-minimization floor rather than unbounded accumulation; this could be probed by training with more capacity or periodic activation functions to see whether the floor drops, which would indicate a capacity limitation rather than a fundamental PINN barrier.
  • My inference: a corrected, fully specified residual system with stated material constants would make this a clean reproduciblity test for any PINN implementation, and the paper's benchmark would be strengthened by reporting residuals at the exact solution rather than only the learned solution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper trains a fully connected PINN with hard-constrained output transformations on a one-dimensional coupled electro-elastodynamic system for linear piezoelectricity. It uses a three-stage optimizer (Adam, AdamW, L-BFGS) and reports global relative L2 errors of 2.34% for displacement and 4.87% for electric potential against a claimed exact standing-wave solution. The central claim is that this validates PINNs as mesh-free solvers for coupled time-dependent PDE systems.

Significance. If the validation were sound, the paper would provide a useful demonstration of PINNs on a coupled multiphysics benchmark and a transparent account of error accumulation. The detailed architecture description and the three-stage optimization strategy are potentially reproducible elements. However, the validity of the entire empirical contribution rests on the claimed exact solution (11)-(12). As shown below, that solution does not satisfy the residuals actually minimized in (18)-(19) for any stated or physically stated parameters, and the material parameters are never specified. The reported L2 errors therefore do not measure the network's ability to solve the coupled PDE system. The manuscript does not currently provide a valid benchmark or a sound validation of the central claim.

major comments (3)
  1. [II-A, Eqs. (11)-(12) and II-C, Eqs. (18)-(19)] The claimed exact solution u = sin(pi x) cos(pi t), phi = 0.5 sin(pi x) cos(pi t) is not a solution of the residuals as written. Direct substitution gives r1 = pi^2 (c^E - rho - 0.5 e33) sin(pi x) cos(pi t), and r2 = -pi^2 (epsilon0 + e33) sin(pi x) cos(pi t) (with epsilon0 appearing twice). For these to vanish one would need rho = c^E - 0.5 e33 and e33 = -epsilon0. The manuscript gives no material parameters and does not state these relations. Moreover, Eq. (19) uses epsilon0 phi_xx, whereas applying Gauss's law (8) to the constitutive relation (7) gives epsilon^S phi_xx, so the residual as written does not even correspond to the governing PDE system. If Eq. (19) is corrected to epsilon^S, the second condition becomes e33 = -0.5(epsilon0 + epsilon^S), again unstated. The reported errors therefore compare against a target that is not a solution of the trained residuals.
  2. [II-C, Eq. (23) and II-B, Eqs. (14)-(17)] The initial-condition loss (23) uses the exact solution u_exact, phi_exact, and the hard-constraint ansatz (14)-(17) already contains sin(pi x)(1-t) and 0.5 sin(pi x)(1-t), which match the initial profile at t=0. The tiny initial-time errors in Fig. 2 are thus by construction: the network is explicitly trained to match the benchmark target at t=0. This does not by itself invalidate the method, but it means the reported global L2 errors conflate interpolation of prescribed initial data with solving the PDE residuals. The paper should report the PDE residual norm (L_PDE) separately and show its convergence; without that, the good agreement at early times cannot be attributed to the physics-informed part of the loss.
  3. [II-A, after Eq. (8); III-A, Eq. (26)] The material parameters rho, c^E, e33, epsilon^S, and epsilon0 are introduced but never assigned numerical values. The residuals (18)-(19) cannot be evaluated, and the reported relative L2 errors (26) cannot be reproduced. If the authors used normalized or dimensionless parameters, that scaling must be stated. If the closed-form solution exists only under special parameter relations, those relations must be given and physically justified. As written, the benchmark is underdetermined.
minor comments (6)
  1. [Abstract and II-C] The abstract describes the PDE constraints as 'soft constraints,' while Section II-C uses hard-constraint basis-function enforcement for boundary/initial conditions. Please reconcile the terminology.
  2. [II-C, Eq. (19)] As noted in Major Comment 1, the coefficient of phi_xx should be epsilon^S (not epsilon0) to match Eq. (8). This is a typo, but it is load-bearing because it enters the residual that is minimized.
  3. [II-B, Eqs. (14)-(17)] The term sin(pi x)(1-t) 'decays away' as t increases, but at t=1 it vanishes, while the exact solution is nonzero there. The network can compensate through u_raw, but the description of this term as encoding the initial condition is misleading; it imposes a time-dependent spatial profile that is not part of the physics.
  4. [III-A, Fig. 1] The y-label of Fig. 1(b) reads 'Electric potential E(x),' but the dependent variable is phi(x,t), not E. Please correct the label.
  5. [Throughout] There are minor grammatical errors (e.g., 'inputed' in II-E, 'primarily' in II-H, 'V oigt' in II-A). A careful proofread would improve readability.
  6. [II-F and III] No code or data availability statement is provided, and no convergence history of the PDE residual is shown. For reproducibility, please include the final PDE loss and, ideally, the code or a repository link.

Circularity Check

0 steps flagged

No circularity: the validation benchmark is independent of the training objective, though the exact-solution benchmark itself is questionable on correctness grounds.

full rationale

The paper's derivation chain is: define PDE residuals (18)-(19) from the constitutive and balance laws, enforce boundary/initial data through hard constraints (14)-(17), minimize combined loss (20), and measure global relative L2 error against the closed-form solution (11)-(12). The closed-form solution enters the loss only in the initial-condition term L_IC (Eqs. 23-24), which is standard problem data, and in the hard-constraint initial profile sin(pi x)(1-t); it is not used to define the interior PDE residual L_PDE. Thus the network's temporal evolution is not forced to match the exact solution by construction; it must emerge from minimizing the PDE residuals. No parameter is fitted to the exact solution and then reported as a prediction. There are no self-citations used as load-bearing uniqueness or ansatz justifications; citations to prior PINN literature (Raissi et al., Cuomo et al., Sukumar & Srivastava) are standard methodological references and are not the present authors' own prior work. The concern that the closed-form solution (11)-(12) does not actually satisfy residuals (18)-(19) as written (e.g., Eq. 19 has an epsilon0/epsilon^S coefficient issue and requires unstated parameter relations such as e33 = -epsilon0) is a substantive correctness/typo problem with the benchmark, not a circularity: the network is still being asked to minimize a fixed residual, and the reported accuracy measures distance to a proposed reference solution. The manuscript's own stated limitations (error accumulation over time, electric-field error amplification) are consistent with an honest benchmark rather than a circular one. Therefore no step in the paper's reasoning reduces by definition or by self-citation to its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The central undefined inputs are the material coefficients; they are neither given nor derivable from the stated exact solution unless special unphysical choices are made. The exact solution itself is an unverified assumption, and the hard-constraint ansatz is an ad hoc addition.

free parameters (1)
  • PDE material coefficients (ρ, cE, e33, εS, ε0) = unspecified
    These must appear in the residual losses (18)–(19). Without numeric values the training is underspecified; the stated 'exact solution' only satisfies the residuals under hidden constraints (e.g., e33 = -ε0 in Eq 19).
axioms (3)
  • domain assumption Linear piezoelectric stress-charge model in 1D is an accurate description of the physical system
    The paper adopts Eq (5)–(8) from refs [13]–[15] without derivation; if the model is wrong, the whole benchmark is wrong.
  • ad hoc to paper Eqs (11)–(12) are an exact solution of the PDE system
    The paper states this in Sec II-A but plugging into the residuals gives nonzero r1/r2 unless hidden constraints on material parameters hold.
  • ad hoc to paper The hard-constraint term sin(πx)(1-t) in Eqs (14)–(17) is a valid ansatz that does not prevent the network from learning the true time evolution
    This term is zero at boundaries and matches the initial condition at t=0, but it is not a solution to the PDEs and adds a forcing bias that the network must compensate.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Unified Physics-Informed Neural Network for Modeling Coupled Electro- and Elastodynamic Wave Propagation Using Three-Stage Loss Optimization." pith.science (2026). https://pith.science/paper/JEECY46B

@misc{pith2026260213811,
  author       = {Pith},
  title        = {Pith review of: A Unified Physics-Informed Neural Network for Modeling Coupled Electro- and Elastodynamic Wave Propagation Using Three-Stage Loss Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEECY46B}},
  note         = {Machine review of arXiv:2602.13811}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Physics-Informed Neural Networks present a novel approach in SciML that integrates physical laws in the form of partial differential equations directly into the NN through soft constraints in the loss function. This work studies the application of PINNs to solve a one dimensional coupled electro-elastodynamic system modeling linear piezoelectricity in stress-charge form, governed by elastodynamic and electrodynamic equations. Our simulation employs a feedforward architecture, mapping space-time coordinates to mechanical displacement and electric potential. Our PINN model achieved global relative L2 errors of 2.34 and 4.87 percent for displacement and electric potential respectively. The results validate PINNs as effective mesh free solvers for coupled time-dependent PDE systems, though challenges remain regarding error accumulation and stiffness in coupled eigenvalue systems.

Figures

Figures reproduced from arXiv: 2602.13811 by Reuben Thomas Thovelil, Suhas Suresh Bharadwaj.

Figure 1
Figure 1. Figure 1: (a) Elastic wave displacement u(x) with respect to position; (b) Electric potential E(x) with respect to position [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Absolute displacement error; (b) Absolute Electric potential error [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

24 extracted references · 2 linked inside Pith

  1. [1]

    Scientific machine learning through physics–informed neural networks: Where we are and what’s next,

    S. Cuomo, V . S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli, “Scientific machine learning through physics–informed neural networks: Where we are and what’s next,”Journal of Scientific Computing, vol. 92, no. 3, p. 88, 2022

    2022

  2. [2]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics, vol. 378, pp. 686–707, 2019

    2019

  3. [3]

    Exploring physics-informed neural networks: From fundamentals to applications in complex systems,

    S. Ganga and Z. Uddin, “Exploring physics-informed neural networks: From fundamentals to applications in complex systems,”arXiv preprint arXiv:2410.00422, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    M-pinn: A mesh-based physics-informed neural network for linear elastic problems in solid mechanics,

    L. Wang, G. Liu, G. Wang, and K. Zhang, “M-pinn: A mesh-based physics-informed neural network for linear elastic problems in solid mechanics,”International journal for numerical methods in engineering, vol. 125, no. 9, p. e7444, 2024

    2024

  5. [5]

    Physics-informed neural network for solving a one- dimensional solid mechanics problem,

    V . Singh, D. Harursampath, S. Dhawan, M. Sahni, S. Saxena, and R. Mallick, “Physics-informed neural network for solving a one- dimensional solid mechanics problem,”Modelling, vol. 5, no. 4, pp. 1532–1549, 2024

    2024

  6. [6]

    Physics- informed neural networks for heat transfer problems,

    S. Cai, Z. Wang, S. Wang, P. Perdikaris, and G. E. Karniadakis, “Physics- informed neural networks for heat transfer problems,”Journal of Heat Transfer, vol. 143, no. 6, p. 060801, 2021

    2021

  7. [7]

    Physics-informed neural network simulation of thermal cavity flow,

    E. Fowler, C. J. McDevitt, and S. Roy, “Physics-informed neural network simulation of thermal cavity flow,”Scientific Reports, vol. 14, no. 1, p. 15203, 2024

    2024

  8. [8]

    Physics- informed neural networks (pinns) for fluid mechanics: A review,

    S. Cai, Z. Mao, Z. Wang, M. Yin, and G. E. Karniadakis, “Physics- informed neural networks (pinns) for fluid mechanics: A review,”Acta Mechanica Sinica, vol. 37, no. 12, pp. 1727–1738, 2021

    2021

  9. [9]

    A review of physics- informed machine learning in fluid mechanics,

    P. Sharma, W. T. Chung, B. Akoush, and M. Ihme, “A review of physics- informed machine learning in fluid mechanics,”Energies, vol. 16, no. 5, p. 2343, 2023

    2023

  10. [10]

    Multiscale physics-informed neural networks for stiff chemical kinetics,

    Y . Weng and D. Zhou, “Multiscale physics-informed neural networks for stiff chemical kinetics,”The Journal of Physical Chemistry A, vol. 126, no. 45, pp. 8534–8543, 2022

    2022

  11. [11]

    Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems,

    A. D. Jagtap, E. Kharazmi, and G. E. Karniadakis, “Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems,”Computer Methods in Applied Mechanics and Engineering, vol. 365, p. 113028, 2020

    2020

  12. [12]

    Solving forward and in- verse problems of contact mechanics using physics-informed neural networks,

    T. Sahin, M. von Danwitz, and A. Popp, “Solving forward and in- verse problems of contact mechanics using physics-informed neural networks,”Advanced Modeling and Simulation in Engineering Sciences, vol. 11, no. 1, p. 11, 2024

    2024

  13. [13]

    A problem in control of elastodynamics with piezo- electric effects,

    H. Antilet al., “A problem in control of elastodynamics with piezo- electric effects,”IMA Journal of Numerical Analysis, vol. 40, no. 4, pp. 2839–2876, 2020

    2020

  14. [14]

    On well-posedness for a piezo-electromagnetic cou- pling model,

    R. Picardet al., “On well-posedness for a piezo-electromagnetic cou- pling model,”arXiv preprint arXiv:1704.01816, 2017

    Pith/arXiv arXiv 2017

  15. [15]

    Institution of electrical and electrical engineers’ standard on piezoelec- tricity,

    “Institution of electrical and electrical engineers’ standard on piezoelec- tricity,” IEEE, iEEE Std 176-1987 (R1997)

    1987

  16. [16]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear pdes,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear pdes,”Journal of Computational Physics, vol. 378, pp. 686–707, 2019

    2019

  17. [17]

    Scientific machine learning through physics-informed neural networks: Where we are and what’s next,

    S. Cuomoet al., “Scientific machine learning through physics-informed neural networks: Where we are and what’s next,”Journal of Scientific Computing, vol. 92, no. 3, p. 88, 2022

    2022

  18. [18]

    Solving high-dimensional partial differential equations using deep learning,

    J. Han, A. Jentzen, and E. Weinan, “Solving high-dimensional partial differential equations using deep learning,”Proceedings of the National Academy of Sciences USA, vol. 115, no. 34, pp. 8683–8688, 2018

    2018

  19. [19]

    Solving real-world optimization tasks using physics-informed neural computing,

    J. Seo, “Solving real-world optimization tasks using physics-informed neural computing,”Scientific Reports, vol. 14, no. 1, p. 202, 2024

    2024

  20. [20]

    Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks,

    N. Sukumar and A. Srivastava, “Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks,” Computer Methods in Applied Mechanics and Engineering, vol. 389, p. 114333, 2022

    2022

  21. [21]

    Three ways to solve partial differential equations with neural networks—a review,

    J. Blechschmidt and O. G. Ernst, “Three ways to solve partial differential equations with neural networks—a review,”Gamm-mitteilungen, vol. 44, no. 2, p. e202100006, 2021

    2021

  22. [22]

    The old and the new: Can physics-informed deep-learning replace traditional linear solvers?

    S. Markidis, “The old and the new: Can physics-informed deep-learning replace traditional linear solvers?”Frontiers in big Data, vol. 4, p. 669097, 2021

    2021

  23. [23]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks,

    S. Wang, Y . Teng, and P. Perdikaris, “Understanding and mitigating gradient flow pathologies in physics-informed neural networks,”SIAM Journal on Scientific Computing, vol. 43, no. 5, pp. A3055–A3081, 2021

    2021

  24. [24]

    Artificial neural networks for solving ordinary and partial differential equations,

    I. E. Lagaris, A. Likas, and D. I. Fotiadis, “Artificial neural networks for solving ordinary and partial differential equations,”IEEE transactions on neural networks, vol. 9, no. 5, pp. 987–1000, 1998

    1998

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.