pith. sign in

arxiv: 2606.03210 · v1 · pith:N27V5624new · submitted 2026-06-02 · 💻 cs.CE · cs.LG· cs.NA· math.NA

Critical evaluation of PINN for FWD inverse analysis and differentiable FEM as an alternative

Yongjin Choi , Hyeonbin Moon , Seunghwa Ryu This is my paper

Pith reviewed 2026-06-28 08:19 UTC · model grok-4.3

classification 💻 cs.CE cs.LGcs.NAmath.NA
keywords PINNDifferentiable FEMFWD backcalculationinverse analysispavement moduliphysics-informed neural networkslayered systems
0
0 comments X

The pith

DiffFEM recovers layer moduli more accurately and stably than PINN or XPINN for multilayer pavement backcalculation.

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

The paper tests automatic-differentiation methods for inverting falling weight deflectometer measurements to recover pavement layer properties. A standard physics-informed neural network fails to recover the moduli because of abrupt changes between layers, while an extended version with domain decomposition still depends heavily on loss weighting and network design and loses accuracy when noise is added. A differentiable finite element approach produces higher accuracy, greater stability, and lower computation time by solving the governing equations exactly at each step instead of penalizing mismatches inside a loss function. The comparison is performed on synthetic layered-pavement data.

Core claim

On a synthetic benchmark of multilayer pavement systems, DiffFEM recovers the layer moduli with higher accuracy and stability than both standard PINN and XPINN, while requiring less computational effort; the advantage stems from enforcing the governing differential equations directly in the solver rather than penalizing deviations through weighted loss terms.

What carries the argument

Differentiable finite element method (DiffFEM) that treats the equilibrium equations as hard constraints satisfied by the forward solver at every iteration.

If this is right

  • DiffFEM inversion remains accurate under added measurement noise where PINN performance drops sharply.
  • XPINN requires extensive tuning of loss weights and architecture to approach DiffFEM accuracy.
  • Computational cost is lower for DiffFEM because no neural network training is needed once the differentiable solver exists.
  • Sharp material discontinuities between layers are handled without special loss balancing.

Where Pith is reading between the lines

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

  • When an efficient differentiable forward solver can be written for a given physics problem, it may be the default choice for gradient-based inversion rather than a neural-network surrogate.
  • The same hard-constraint strategy could be applied to other inverse problems that involve layered or discontinuous domains, such as composite structures or subsurface imaging.
  • Hybrid schemes that embed a differentiable solver inside a neural network might combine the strengths of both approaches if pure DiffFEM is unavailable.
  • Validation against actual field data with known ground truth is required before claiming practical superiority for engineering use.

Load-bearing premise

The synthetic benchmark data and chosen network architectures are representative of the difficulties that appear in real falling weight deflectometer measurements.

What would settle it

A direct comparison of the same inversion pipelines on a collection of real FWD field records whose layer moduli have been independently measured would show whether the accuracy and stability advantage persists.

Figures

Figures reproduced from arXiv: 2606.03210 by Hyeonbin Moon, Seunghwa Ryu, Yongjin Choi.

Figure 1
Figure 1. Figure 1: Schematic of the falling weight deflectometer (FWD) test on a layered pavement system. An impulse load [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Computational flow of the DiffFEM inversion framework. The layer moduli [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the XPINN inversion framework. Three independent sub-networks [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Axisymmetric finite element domain for the three-layer pavement system. The mesh spans [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Surface deflection basins measured by FWD geophone sensors under three noise conditions: noise-free, [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Convergence histories for the noise-free inversion. Top row: XPINN loss components (left) and inferred [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Predicted displacement fields and pointwise errors for the noise-free case. Left columns: ground-truth [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Modulus convergence history for the vanilla PINN formulation without noise. Dashed horizontal lines [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Convergence histories for the noisy inversion with [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Predicted displacement fields and pointwise errors for [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Convergence histories for the noisy inversion with [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Predicted displacement fields and pointwise errors for [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Summary of inversion accuracy across all noise levels ( [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
read the original abstract

Automatic-differentiation-based inverse analysis methods, including physics-informed neural networks (PINNs) and differentiable programming, have recently shown great promise due to their ability to compute accurate gradients and convergence efficiency. However, their applicability to falling weight deflectometer (FWD) backcalculation remains unexplored. This study critically evaluates PINN-based inverse analysis for a multilayer pavement system and investigates differentiable finite element method (DiffFEM) as an alternative based on a synthetic benchmark. The standard PINN does not recover layer moduli because of the sharp domain discontinuities inherent to layered pavement systems. Although we use an extended PINN with domain decomposition (XPINN), which shows better performance on discontinuous domains, its performance remains highly sensitive to loss weighting and network architecture, and degrades under measurement noise. By contrast, DiffFEM consistently achieves more accurate, stable, and computationally efficient inversion results. These results indicate that DiffFEM, which enforces the governing physics as a hard constraint, yields better accuracy, robustness, and computational efficiency than PINN-based approaches, in which the governing physics is imposed as a soft constraint through the loss function. More broadly, the findings suggest that the choice between PINN- and DiffFEM-based inverse analysis needs careful consideration, with DiffFEM offering practical advantages when an efficient and robust differentiable forward solver is available.

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

3 major / 2 minor

Summary. The paper claims that standard PINNs fail to recover layer moduli in multilayer pavement FWD inverse analysis due to sharp domain discontinuities, while XPINNs improve but remain highly sensitive to loss weights, architecture, and degrade under noise; by contrast, DiffFEM yields more accurate, stable, and efficient results on a synthetic benchmark because it enforces the governing PDEs as a hard constraint rather than a soft loss term.

Significance. If the empirical comparison holds, the work supplies concrete evidence that hard-constraint differentiable solvers can outperform soft-constraint neural approaches on discontinuous domains typical of civil-engineering inverse problems, offering practical selection criteria when an efficient differentiable forward model exists.

major comments (3)
  1. [§3 and §4] §3 (Methods) and §4 (Results): the synthetic benchmark generation procedure, including the precise noise model, layer-property sampling ranges, and mesh discretization, is not specified with sufficient detail or equations to permit reproduction or independent verification of the reported sensitivity and accuracy differences.
  2. [§4.2] §4.2 (Quantitative comparison): no tables or figures report explicit error metrics (e.g., relative L2 error on recovered moduli E1–E4 or deflection residuals) across noise levels for DiffFEM versus XPINN; only qualitative statements are given, which is load-bearing for the central claim of consistent superiority.
  3. [§4.3] §4.3 (XPINN sensitivity): the statement that performance “remains highly sensitive to loss weighting and network architecture” lacks the specific weight schedules, architecture variants, or ablation results that would quantify the performance variation and support the robustness contrast with DiffFEM.
minor comments (2)
  1. [Figures 4–7] Figure captions and axis labels in the results section should explicitly state the noise standard deviation and number of Monte-Carlo realizations used for each curve.
  2. [Abstract] The abstract’s final sentence on “practical advantages when an efficient and robust differentiable forward solver is available” would benefit from a short qualifying clause referencing the computational cost of assembling the differentiable stiffness matrix.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. These points identify key areas where the manuscript can be strengthened for reproducibility and quantitative rigor. We address each major comment below and will incorporate the suggested additions in the revised version.

read point-by-point responses

  1. Referee: [§3 and §4] §3 (Methods) and §4 (Results): the synthetic benchmark generation procedure, including the precise noise model, layer-property sampling ranges, and mesh discretization, is not specified with sufficient detail or equations to permit reproduction or independent verification of the reported sensitivity and accuracy differences.

    Authors: We agree that the benchmark generation details require expansion for full reproducibility. In the revised manuscript we will add a dedicated subsection to §3 that specifies: the exact sampling ranges used for each layer modulus (E1: 80–600 MPa, E2: 40–300 MPa, E3: 20–150 MPa, E4: 10–80 MPa), the noise model (zero-mean Gaussian noise scaled to 0 %, 5 %, and 10 % of the peak deflection), and the mesh parameters (uniform 0.1 m element size, 20 elements per layer). The forward-model equation used to generate the synthetic FWD deflections will also be stated explicitly. revision: yes

  2. Referee: [§4.2] §4.2 (Quantitative comparison): no tables or figures report explicit error metrics (e.g., relative L2 error on recovered moduli E1–E4 or deflection residuals) across noise levels for DiffFEM versus XPINN; only qualitative statements are given, which is load-bearing for the central claim of consistent superiority.

    Authors: We accept that explicit numerical metrics are needed to substantiate the superiority claim. The revised §4.2 will contain a new table that lists, for each method and each noise level, the relative L2 error on E1–E4 and the L2 norm of the deflection residual. These values will be computed from the same synthetic cases already used in the study. revision: yes

  3. Referee: [§4.3] §4.3 (XPINN sensitivity): the statement that performance “remains highly sensitive to loss weighting and network architecture” lacks the specific weight schedules, architecture variants, or ablation results that would quantify the performance variation and support the robustness contrast with DiffFEM.

    Authors: We agree that the sensitivity statement should be supported by concrete ablation data. In the revised §4.3 we will add an ablation table showing recovered-modulus errors for XPINN under five different PDE-loss weight schedules (1, 10, 50, 100, 500) and three network depths (4, 6, 8 layers). The same table will report the corresponding errors obtained with DiffFEM, thereby quantifying the robustness contrast. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical comparison on synthetic benchmarks

full rationale

The paper conducts an empirical evaluation of PINN/XPINN versus DiffFEM on synthetic FWD inverse problems, comparing accuracy, stability, and efficiency under hard versus soft physics enforcement. No derivation chain exists that reduces a claimed prediction or result to its own inputs by construction, self-definition, or self-citation load-bearing steps. The central findings rest on numerical experiments whose outcomes are not forced by the method definitions themselves, and the work explicitly limits conclusions to the benchmark results without invoking uniqueness theorems or ansatzes from prior self-work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review; the paper appears to rest on standard assumptions of linear elasticity for pavement layers, differentiability of the FEM solver, and that the chosen synthetic load cases capture the essential discontinuity effects. No explicit free parameters or invented entities are named in the abstract.

axioms (2)
  • domain assumption Linear elastic constitutive model for each pavement layer
    Implicit in any FWD back-calculation that treats layers as homogeneous elastic media; invoked when defining the forward problem for both PINN and DiffFEM.
  • domain assumption The differentiable FEM implementation correctly computes exact gradients of the forward map
    Required for DiffFEM to serve as a hard-constraint solver; location would be in the methods section describing the automatic-differentiation pipeline.

pith-pipeline@v0.9.1-grok · 5776 in / 1619 out tokens · 32515 ms · 2026-06-28T08:19:35.063740+00:00 · methodology

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

51 extracted references · 3 linked inside Pith

  1. [1]

    Computer-Aided Civil and Infrastructure Engineering , volume=

    Prediction of stratified ground consolidation via a physics-informed neural network utilizing short-term excess pore water pressure monitoring data , author=. Computer-Aided Civil and Infrastructure Engineering , volume=. 2025 , publisher=

    2025

  2. [2]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Physics-informed neural network-based discovery of hyperelastic constitutive models from extremely scarce data , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2025 , publisher=

    2025

  3. [3]

    arXiv preprint arXiv:2602.05849 , year=

    Visualizing the loss landscapes of physics-informed neural networks , author=. arXiv preprint arXiv:2602.05849 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2402.01868 , year=

    Challenges in training pinns: A loss landscape perspective , author=. arXiv preprint arXiv:2402.01868 , year=

    arXiv

  5. [5]

    Journal of Scientific Computing , volume=

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next , author=. Journal of Scientific Computing , volume=. 2022 , publisher=

    2022

  6. [6]

    Advanced Modeling and Simulation in Engineering Sciences , volume=

    Solving forward and inverse problems of contact mechanics using physics-informed neural networks , author=. Advanced Modeling and Simulation in Engineering Sciences , volume=. 2024 , publisher=

    2024

  7. [7]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Energy-based physics-informed neural network for frictionless contact problems under large deformation , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2025 , publisher=

    2025

  8. [8]

    Nonlinear Dynamics , volume=

    Physics-informed neural networks for friction-involved nonsmooth dynamics problems , author=. Nonlinear Dynamics , volume=. 2024 , publisher=

    2024

  9. [9]

    Computers and Geotechnics , volume=

    A physics-informed deep neural network for surrogate modeling in classical elasto-plasticity , author=. Computers and Geotechnics , volume=. 2023 , publisher=

    2023

  10. [10]

    Journal of Hydrology , volume=

    Physics-informed neural network for solution of forward and inverse kinematic wave problems , author=. Journal of Hydrology , volume=. 2024 , publisher=

    2024

  11. [11]

    SIAM Journal on Scientific Computing , volume=

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks , author=. SIAM Journal on Scientific Computing , volume=. 2021 , publisher=

    2021

  12. [12]

    Journal of Computational Physics , volume=

    When and why PINNs fail to train: A neural tangent kernel perspective , author=. Journal of Computational Physics , volume=. 2022 , publisher=

    2022

  13. [13]

    Journal of Computational Physics , volume=

    Self-adaptive physics-informed neural networks , author=. Journal of Computational Physics , volume=. 2023 , publisher=

    2023

  14. [14]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Multi-objective loss balancing for physics-informed deep learning , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2025 , publisher=

    2025

  15. [15]

    arXiv preprint arXiv:2502.07209 , year=

    Enhancing Physics-Informed Neural Networks Through Feature Engineering , author=. arXiv preprint arXiv:2502.07209 , year=

    arXiv

  16. [16]

    Computers and Geotechnics , volume=

    Leveraging physics-informed neural networks in geotechnical earthquake engineering: An assessment on seismic site response analyses , author=. Computers and Geotechnics , volume=. 2025 , publisher=

    2025

  17. [17]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2022 , publisher=

    2022

  18. [18]

    Journal of the Chinese Institute of Engineers , volume=

    Development of 3-D finite element model for flexible pavements , author=. Journal of the Chinese Institute of Engineers , volume=. 2004 , publisher=

    2004

  19. [19]

    arXiv preprint arXiv:2603.08824 , year=

    SoftJAX & SoftTorch: Empowering Automatic Differentiation Libraries with Informative Gradients , author=. arXiv preprint arXiv:2603.08824 , year=

    arXiv

  20. [20]

    ACM SIGGRAPH 2007 papers , pages=

    Efficient symbolic differentiation for graphics applications , author=. ACM SIGGRAPH 2007 papers , pages=

    2007

  21. [21]

    arXiv preprint arXiv:1907.07587 , year=

    A differentiable programming system to bridge machine learning and scientific computing , author=. arXiv preprint arXiv:1907.07587 , year=

    Pith/arXiv arXiv 1907

  22. [22]

    arXiv preprint arXiv:2509.21393 , year=

    Impact of loss weight and model complexity on physics-informed neural networks for computational fluid dynamics , author=. arXiv preprint arXiv:2509.21393 , year=

    arXiv

  23. [23]

    International Journal of Geomechanics , volume=

    Modeling of the FWD deflection basin to evaluate airport pavements , author=. International Journal of Geomechanics , volume=. 2014 , publisher=

    2014

  24. [24]

    arXiv preprint arXiv:1704.08863 , year=

    On weight initialization in deep neural networks , author=. arXiv preprint arXiv:1704.08863 , year=

    Pith/arXiv arXiv

  25. [25]

    Mathematical programming , volume=

    On the limited memory BFGS method for large scale optimization , author=. Mathematical programming , volume=. 1989 , publisher=

    1989

  26. [26]

    IEEE Robotics and Automation Letters , volume=

    Sim-to-real for soft robots using differentiable fem: Recipes for meshing, damping, and actuation , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  27. [27]

    Computer Physics Communications , volume=

    JAX-FEM: A differentiable GPU-accelerated 3D finite element solver for automatic inverse design and mechanistic data science , author=. Computer Physics Communications , volume=. 2023 , publisher=

    2023

  28. [28]

    arXiv preprint arXiv:1910.00935 , year=

    Difftaichi: Differentiable programming for physical simulation , author=. arXiv preprint arXiv:1910.00935 , year=

    arXiv 1910

  29. [29]

    arXiv preprint arXiv:2512.24365 , year=

    Deep Learning in Geotechnical Engineering: A Critical Assessment of PINNs and Operator Learning , author=. arXiv preprint arXiv:2512.24365 , year=

    Pith/arXiv arXiv

  30. [30]

    Computers & Industrial Engineering , pages=

    Fundamental flaws of physics-informed neural networks and explainability methods in engineering systems , author=. Computers & Industrial Engineering , pages=. 2025 , publisher=

    2025

  31. [31]

    Computers and Geotechnics , volume=

    Inverse analysis of granular flows using differentiable graph neural network simulator , author=. Computers and Geotechnics , volume=. 2024 , publisher=

    2024

  32. [32]

    Automatic differentiation in pytorch , author=

  33. [33]

    Journal of Computing and Information Science in Engineering , volume=

    Physics-guided, physics-informed, and physics-encoded neural networks and operators in scientific computing: Fluid and solid mechanics , author=. Journal of Computing and Information Science in Engineering , volume=. 2024 , publisher=

    2024

  34. [34]

    Journal of Geophysical Research: Solid Earth , volume=

    Physics-informed neural networks (PINNs) for wave propagation and full waveform inversions , author=. Journal of Geophysical Research: Solid Earth , volume=. 2022 , publisher=

    2022

  35. [35]

    Geomechanics and Geoengineering , volume=

    Physics-informed machine learning in geotechnical engineering: a direction paper , author=. Geomechanics and Geoengineering , volume=. 2025 , publisher=

    2025

  36. [36]

    Journal of Rock Mechanics and Geotechnical Engineering , volume=

    Physics-informed deep learning for one-dimensional consolidation , author=. Journal of Rock Mechanics and Geotechnical Engineering , volume=. 2021 , publisher=

    2021

  37. [37]

    Wiley interdisciplinary reviews: data mining and knowledge discovery , volume=

    A review of automatic differentiation and its efficient implementation , author=. Wiley interdisciplinary reviews: data mining and knowledge discovery , volume=. 2019 , publisher=

    2019

  38. [38]

    Transportation research record , number=

    Backcalculation of flexible pavement moduli using artificial neural networks , author=. Transportation research record , number=

  39. [39]

    Transportation Research Record , volume=

    Modified Newton algorithm for backcalculation of pavement layer properties , author=. Transportation Research Record , volume=. 1993 , publisher=

    1993

  40. [40]

    Texas Dept

    Development of a new methodology for characterizing pavement structural condition for network-level applications , author=. Texas Dept. of Transportation, Austin, TX , year=

  41. [41]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  42. [42]

    1990 , month = nov, number =

    Modulus 4.0: Expansion and Validation of the Modulus Backcalculation System , author =. 1990 , month = nov, number =

    1990

  43. [43]

    2017 , month = mar, number =

    Using Falling Weight Deflectometer Data With Mechanistic-Empirical Design and Analysis, Volume I: Final Report , author =. 2017 , month = mar, number =

    2017

  44. [44]

    Additive Manufacturing , pages=

    Multi-layer thermal history prediction framework for directed energy deposition based on extended physics-informed neural networks (XPINN) , author=. Additive Manufacturing , pages=. 2025 , publisher=

    2025

  45. [45]

    Communications in Computational Physics , volume=

    Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations , author=. Communications in Computational Physics , volume=. 2020 , publisher=

    2020

  46. [46]

    Journal of machine learning research , volume=

    Automatic differentiation in machine learning: a survey , author=. Journal of machine learning research , volume=

  47. [47]

    I , author=

    The general theory of stresses and displacements in layered systems. I , author=. Journal of Applied Physics , volume=. 1945 , publisher=

    1945

  48. [48]

    Volume III: Guidelines for deflection testing, analysis, and interpretation , author=

    Using falling weight deflectometer data with mechanistic-empirical design and analysis. Volume III: Guidelines for deflection testing, analysis, and interpretation , author=. 2016 , institution=

    2016

  49. [49]

    and Rao, Chetana and Irwin, Lynne , title =

    Von Quintus, Harold L. and Rao, Chetana and Irwin, Lynne , title =. 2015 , month = dec, note =

    2015

  50. [50]

    Construction and Building Materials , volume=

    Co-variance matrix adaptation evolution strategy for pavement backcalculation , author=. Construction and Building Materials , volume=. 2010 , publisher=

    2010

  51. [51]

    2007 , edition=

    Numerical Recipes: The Art of Scientific Computing , author=. 2007 , edition=

    2007