Pith. sign in

REVIEW 3 major objections 5 minor 73 references

Finite basis physics-informed neural networks with hard constraints for viscous fluid flow in highly perforated domains

T0 review · 3 major / 5 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 This paper claims that hard-constrained finite basis PINNs—overlapping localized networks with exact boundary encoding—solve Stokes flow in highly perforated domains with relative errors near 1% at 100 perforations and convergence nearly…

desk verdict A solid FBPINN+hard-constraint paper with unstated composite ansatz; the method works numerically, but the exactness claim needs fixing. read the letter →

arxiv 2608.08114 v1 pith:QBMNJJ67 submitted 2026-08-08 physics.flu-dyn cs.LG

Jeeeun Lee , Denis Korolev , Miro Duhovic , Seong Su Kim This is my paper · ORCID
classification physics.flu-dyncs.LG
keywords physics-informedneuralnetworksfinitebasisPINNshardboundaryconstraintsStokesflowperforateddomainsspectralbiasdomaindecompositiondistancefunctions
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

This paper sets out to establish that a neural-network PDE solver can remain accurate for Stokes flow in domains with many circular obstacles, a setting where ordinary PINNs degrade sharply. The proposed combination is finite basis PINNs—domain decomposition into overlapping, normalized subdomains blended by window functions—with hard constraints that encode the no-slip condition on every perforation boundary exactly, removing the boundary penalty term. The paper argues this mitigates spectral bias, reduces gradient stiffness, and makes convergence only weakly dependent on the number of perforations. A sympathetic reader would care because flow around dense fibre-like obstructions is exactly the multi-scale, high-frequency regime where soft-penalty PINNs struggle, and applications such as resin flow through fibre bundles require many holes. The numerical evidence includes relative $L^2$ errors below 1% for random arrangements of 100 perforations after 100,000 iterations.

What carries the argument

The load-bearing construction is the hard-constrained FBPINN ansatz $$u_\$\theta$(x)=C_w\left(\sum_{i=1}^N \omega_i(x)\, \mathrm{unnorm}_u\circ C_p^i[NN_i\circ \mathrm{norm}_i](x)\right),$$ where the $\omega_i$ are overlapping cosine partition-of-unity windows and each local network sees coordinates normalized to a parent subdomain. The no-slip constraint uses the aggregated inverse distance function $$l_{\partial\Omega_p}(x)=\tanh\!\left(a\left(\sum_{k=1}^K \phi_{\mathrm{disk}}(x,c_k)^{-m}\right)^{-1/m}\right),$$ which vanishes on every perforation boundary, while $C_w$ is an affine lifting operator that imposes the wall profile. Because $\Delta(l\,\bar u)=l\Delta \bar u+2\nabla l\cdot\nabla \bar u+\bar u\,\Delta l$, the distance function injects the perforation geometry into the Stokes residual as an inductive bias, and the paper studies how the localisation parameters $m,a$ control the curvature $\Delta l_{\partial\Omega_p}$ and hence the stiffness of training.

What would settle it

Run the 100-random-perforation benchmark with the wall lifting function $\bar g$ chosen so that it has nonzero support on one perforation boundary; the paper's compatibility warning predicts degraded accuracy or violated no-slip, so observing clean no-slip plus under-1% error would contradict its stated mechanism. Alternatively, measure the maximum of $|\Delta l_{\partial\Omega_p}|$ and train with $m=12,a=25$: the paper predicts errors grow because of curvature stiffness, so flat errors there would falsify that part of the claim.

Watch

Extended reading notes

Core claim

The central discovery is that soft boundary enforcement is the main bottleneck: once no-slip is imposed by multiplying the network output by a smoothed distance-to-perforations function, and the approximation is localized through the FBPINN ansatz, the Stokes residual can be trained to below 1% relative error with dozens to 100 perforations. The paper reports that hard constraints alone are not enough—hard-constrained single-network PINNs still worsen as perforation count grows—and that domain decomposition, input normalization, non-dimensionalisation, and residual-based adaptive collocation must act together. It also provides a Fourier decomposition of the FBPINN ansatz showing two opposing mechanisms: subdomain rescaling lowers the effective frequencies seen by local networks, while window localization broadens the spectral coupling kernel and introduces convolution-induced spectral leakage. The accompanying theory states a universal approximation result for FBPINNs and explains why hard constraints recover stronger $H^1$ velocity and $L^2$ pressure error norms compared with soft penalties.

Load-bearing premise

Everything rests on the assumption that the hand-built distance-to-the-perforations function can be multiplied into the network output and composed with the wall lifting so that both boundary conditions hold exactly while the curvature it injects into the Stokes residual stays mild enough to train; this is not a theorem, and the paper itself warns that wall-lifting support can break compatibility.

Editorial extensions

If this is right

  • Under the paper's protocol, increasing periodic perforations from 16 to 64 does not produce the accuracy collapse seen for soft-constrained or single-network hard-constrained PINNs; the 64-hole case reaches relative $L^2$ errors around half a percent after 100,000 iterations.
  • Hard-constrained FBPINNs also handle random arrangements, with five different 100-perforation geometries all ending below 1% relative $L^2$ error after 100,000 iterations.
  • In the reported 64-perforation test, the FBPINN is faster in wall-clock time than the regular PINN baseline, about 20 versus 42 minutes on an A100 after 100,000 iterations.
  • Increasing the subdomain overlap from 2.0 to 2.4 reduces errors by about 21% but raises compute time by about 58%, so the paper recommends an overlap ratio of 2.0 as a balance.
  • Gradient alignment stays positive early for the hard-constrained FBPINN, whereas the soft-constrained FBPINN remains near $-0.7$, linking hard constraints to reduced boundary-induced gradient conflict.

Reading between the lines

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

  • If the mechanism is as described, the method should extend to three-dimensional fibre arrangements and to more than 100 perforations by keeping subdomain size tied to obstacle spacing rather than domain size; the paper does not run such cases.
  • The Fourier analysis suggests a quantitative design rule the authors do not state explicitly: overlap should be large enough that the window spectrum stays narrow relative to the rescaled target frequencies, so one could test whether the optimal overlap ratio tracks the spectral width of the window.
  • A natural stress test is replacing circular holes with sharp-cornered obstacles; the paper's regularity assumptions and its curvature-stiffness argument imply accuracy would degrade, and the architecture's handling of corner singularities is untested.
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

3 major / 5 minor

Summary. The paper proposes hard-constrained finite basis physics-informed neural networks (FBPINNs) for solving the Stokes equations in two-dimensional domains with many circular perforations. The method combines overlapping domain decomposition with FBPINN window functions, distance-function-based hard constraints for the no-slip condition on the perforations and for the wall boundary, non-dimensionalisation, adaptive loss weighting, and residual-based adaptive collocation refinement. The authors provide a universal approximation proof for FBPINNs, a Fourier analysis of the ansatz, and numerical experiments for periodic (16, 36, 64 perforations) and random (36, 64, 100 perforations) arrangements, comparing against soft- and hard-constrained PINNs and against Taylor–Hood finite element reference solutions.

Significance. If the main claims hold, this work provides a practical and scalable PINN framework for Stokes flow in highly perforated microstructures, a problem of direct relevance to composite manufacturing and microfluidics. The paper has clear strengths: the numerical validation uses external finite element references rather than constructed analytic solutions; the 64-periodic sensitivity study reports seed-averaged errors over five random seeds; and the theoretical sections give a proof of universal approximation and an explicit Fourier-space expression for the FBPINN ansatz. These are substantial contributions. However, the central claim that the implemented ansatz enforces both wall and no-slip boundary conditions exactly is not demonstrated, and the presented experiments do not fully support the headline claim of convergence being only weakly affected by the number of perforations, since most error curves are single runs.

major comments (3)
  1. [§3.2, Eqs. (26), (33), (44)] The actual composite hard-constraint ansatz used in all numerical experiments is never written down. The text says that “both the wall and no-slip boundary hard constraints” are applied globally “as in (26)”, but Eq. (26) is only a single-constraint illustration. If the velocity is formed as C_w[l_{∂Ωp} \bar u] with C_w defined in (44), then on ∂Ωp the lifting term g[1 - tanh(a_w l_{∂Ωw})] is nonzero: for the 64-periodic case it is about 3e-4 relative to |g|. If the ordering is reversed, the wall condition fails. Thus no stated formula satisfies both boundary conditions exactly, and the class (8) — and therefore the error estimate (9) — does not apply to the implemented method. The compatibility warning in §3.2 is not resolved in the numerical section. The authors need to state the exact composite formula, verify analytically or numerically that both boundary conditions are enforced, and quantify the boundary residual if it is not exactly zero.
  2. [§4, Figs. 6–8 and Fig. 15] The central convergence curves that support the claim of convergence “only weakly affected by the number of perforations” are single training runs without error bars or seed statistics. Seed-averaged results are reported only for the 64-periodic case (Tables 1–5), not for 16 and 36 periodic perforations or for 36, 64, and 100 random perforations. For a claim that is explicitly about robustness across problem sizes, the headline comparisons should include at least a few seeds per case or an explicit statement that the plotted curves are representative single runs.
  3. [§3.2, Eq. (28) and Fig. 3] The curvature and stiffness analysis of Δl_{∂Ωp} analyzes only the no-slip factor in isolation. The implemented ansatz composes l_{∂Ωp} with the wall operator C_w defined in (44), so the Laplacian of the constrained velocity also contains derivatives of tanh(a_w l_{∂Ωw}) and of the lifting term g[1 - tanh(a_w l_{∂Ωw})]. The claimed stiffness reduction from hard constraints is therefore not demonstrated for the actual formula used in the experiments. This issue is directly connected to the missing composite formula in the first comment; the analysis should be carried out for the full ansatz.
minor comments (5)
  1. [§2.1, Eq. (2)] In Eq. (2), writing the compatibility condition as ∫_{∂Ωw} g·η_w ds = ∫Ω div u dx = 0 is redundant and slightly confusing; the middle equality follows from the divergence theorem and the zero-divergence condition, but the notation suggests an independent condition. Please simplify.
  2. [§3.2, after Eq. (36)] The sentence “we apply both the wall and no-slip boundary hard constraints globally as in (26)” is difficult to interpret because Eq. (26) contains only one distance function and one boundary datum. Please give the explicit global formula and explain how the two constraints are combined.
  3. [Algorithm 1, line 13] The notation “λ_k ← Gradient-based weight scaling (∇θ,ψLλk(θk,ψk), λ_k)” is ambiguous; specify whether all weights (λ_div, λ_b) are updated together and how the moving average of Eq. (11) is applied.
  4. [Figure 14] The caption of Fig. 14 does not identify which curves in panels (b) and (c) correspond to which model variant; please add a legend or refer to the colors/line styles described in the text.
  5. [Data availability] The data availability statement says code will be made available upon publication; for a numerical methods paper, providing the code or a public repository at submission would strengthen reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claims are validated against external Taylor-Hood FEM references, and no target-solution-dependent parameters are fitted; self-citations are contextual rather than load-bearing.

full rationale

The paper's central results are empirical and are checked against externally computed Taylor-Hood FEM reference solutions, not against quantities derived from the neural-network ansatz itself or from fitted parameters. The hard-constraint parameters (a_w=10, m=2, a=5) are hand-chosen hyperparameters, and their influence is studied in sensitivity tables, so no 'prediction' reduces to a fit. The FBPINN structure is adopted from the external reference [46], the distance-function construction from [35], and the approximation-theoretic bounds from [49]; these are building blocks, and the paper does not rely on a self-citation chain to establish its main convergence or accuracy claims. The self-citations [4], [5], and [7] are contextual and do not carry the load of the argument. One non-circular limitation should be flagged: in Section 3.2 the paper states that it 'apply both the wall and no-slip boundary hard constraints globally as in (26),' and in Section 4 defines the wall operator (44), but it never writes the exact composite formula that would simultaneously enforce u=g on ∂Ω_w and u=0 on ∂Ω_p. As written, the lifting term g[1-tanh(a_w l_∂Ωw)] is generally nonzero on the perforation boundaries, so exact no-slip enforcement is not demonstrated. This is a completeness and correctness verification gap, not a circular reduction: the method is not defined in terms of its own predictions, and the numerical accuracy comparisons remain meaningful against external FEM solutions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard approximation-theoretic facts (density of tanh networks, loss-error equivalence from [49]), on regularity of the Stokes solution in the chosen domains, and on hand-chosen hyperparameters controlling the hard constraints. The paper introduces no new physical entities or fitted physical constants; the method's parameters are training hyperparameters whose sensitivity is partially characterized in Tables 1-5.

free parameters (4)
  • tanh modulation gain a = 5 (all cases)
    Chosen by hand; controls localization of the no-slip distance function. Sensitivity analysis (Tables 2-3) shows errors increase for larger a.
  • aggregation exponent m = 2 (all cases)
    Chosen by hand; controls how strongly l_inv is dominated by the nearest perforation. Larger m increases curvature of Δl_∂Ωp and degrades accuracy.
  • wall modulation gain a_w = 10
    Chosen to make the wall lifting negligible away from the boundary while satisfying the wall data exactly.
  • subdomain overlap ratio δ = 2.0
    Chosen as a balance between stability and cost; Table 1 shows accuracy improves with larger overlap but compute time grows.
assumptions (5)
  • domain assumption The window functions form a partition of unity and each subdomain is Lipschitz (perforation diameter smaller than subdomain side).
    Needed for Proposition 3.1 and the FBPINN ansatz (15).
  • standard math Two-layer tanh networks can approximate H^3 functions in H^2 and H^1 pressure to arbitrary accuracy (density result of [59]).
    Used in Remark 2.1 and Proposition 3.1 to justify approximation error bounds.
  • domain assumption The Stokes solution satisfies u∈H^2(Ω)^d and p∈H^1(Ω) with p mean-zero, and the perforations are mutually separated C^2 curves.
    Required for the loss-to-error estimates (7) and (9) from [49] and for the distance function derivatives to be bounded.
  • standard math The error estimates of [49, Theorem 8, Remark 13] hold for the hard-constrained network class (8).
    Equation (9) is the basis for claiming convergence in H^1 velocity and L^2 pressure norms.
  • domain assumption The Taylor-Hood FEM reference solutions are sufficiently resolved to serve as ground truth.
    All error measurements in Section 4 are computed against these references; mesh details are not provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Finite basis physics-informed neural networks with hard constraints for viscous fluid flow in highly perforated domains." pith.science (2026). https://pith.science/paper/QBMNJJ67

@misc{pith2026260808114,
  author       = {Pith},
  title        = {Pith review of: Finite basis physics-informed neural networks with hard constraints for viscous fluid flow in highly perforated domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBMNJJ67}},
  note         = {Machine review of arXiv:2608.08114}
}
read the original abstract

In this work, viscous fluid flow governed by the Stokes equations in highly perforated domains is studied using physics-informed neural networks (PINNs). Perforated microstructures induce complex boundary conditions and fine-scale flow features that are difficult for standard neural networks to resolve. Conventional PINNs, even when combined with advanced training techniques, can suffer from a loss of accuracy and efficiency as the number of perforations increases. One important source of this difficulty is the soft enforcement of boundary conditions through penalty terms, which can lead to stiffness, gradient conflicts, and poor resolution of near-boundary flow structures. Hard constraints provide an alternative by encoding boundary conditions exactly into the network ansatz, but may introduce undesirable non-local effects due to the global nature of the approximation. To address these challenges, finite basis PINNs (FBPINNs), which are based on domain decomposition and localisation principles, are used together with hard boundary constraints that efficiently encode perforation-related boundary conditions. This approach helps mitigate spectral bias, improves overall accuracy, and exhibits convergence that is only weakly affected by the number of perforations, thereby providing an efficient and highly parallelisable neural network framework. The proposed approach is further supported with theoretical arguments, specifically focusing on the localisation and approximation properties of FBPINNs.

Figures

Figures reproduced from arXiv: 2608.08114 by the authors.

Figure 1
Figure 1. Plots of the velocity magnitude contours and the velocity magnitude at [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Effect of subdomain partition fineness and overlap on windows and their Fourier spectra: shrinking physical support of the window functions by either increasing the number of subdomains N or decreasing the overlap δ broadens their representative Fourier spectrum. The plotted spectra correspond to the Fourier spectra of the windows located at the centres of the respective partitions. in the rescaled form σiξ. Since t… view at source ↗
Figure 3
Figure 3. Localisation and curvature of l∂Ωp (K = 64, R = 0.03). The top row shows l (m) inv for m = 2, 4, 8, 12, while the middle row shows l∂Ωp = tanh(a, l (m) inv ) for (m, a) = (2, 5),(4, 10), (8, 15), (12, 25). The bottom row displays the signed logarithmic transform sign(∆l∂Ωp ) log10(1 + |∆l∂Ωp |) for the same parameter pairs. Larger values of m and a lead to stronger localisation, but also generate sharper curvature p… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Example of the randomly and adaptively-sampled collocation points after 5,000 and 70,000 iterations of training for 16 periodic perfora [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Unit-domain geometry for the 16-perforation case, showing the microstructure region in grey, where the [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the L 2 and microstructure L 1 errors for the soft- and hard-constrained variants of PINNs and FBPINNs for the 16- perforation case. The quantities u, v and p represent the errors in the horizontal velocity, vertical velocity, and pressure, respectively. …
Figure 7
Figure 7. Figure 7: Comparison of the L 2 and microstructure L 1 errors for the soft- and hard-constrained variants of PINNs and FBPINNs for the 36- perforation case. The quantities u, v and p represent the errors in the horizontal velocity, vertical velocity, and pressure, respectively. …
Figure 8
Figure 8. Figure 8: Comparison of the L 2 and microstructure L 1 errors for the soft- and hard-constrained variants of PINNs and FBPINNs for the 64- perforation case. The quantities u, v and p represent the errors in the horizontal velocity, vertical velocity, and pressure, respectively. …
Figure 9
Figure 9. Figure 9: Comparison of the hard-constrained FBPINN predictions and reference solution for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Comparison of the hard-constrained FBPINN predictions and reference solution for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Comparison of the hard-constrained FBPINN predictions and reference solution for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Comparison of the hard-constrained FBPINN predictions and reference solution in the microstructure region for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Comparison of the hard-constrained FBPINN predictions and reference solution in the microstructure region for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 8
Figure 8. Figure 8: Since the alignment score only quantifies directional conflicts, and not the magnitudes of the loss gradients, [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 14
Figure 14. Figure 14: Analysis of the results for the 64-perforation case. a) Histograms of the back-propagated loss gradients after 50,000 iterations for the [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Comparison of the overall L 2 errors and microstructure L 1 errors for the hard-constrained FBPINN model in random arrangements of 36, 64, and 100 perforations. The quantities u, v, and p denote the errors in the horizontal velocity, vertical velocity, and pressure, r…
Figure 16
Figure 16. Figure 16: Comparison of the hard-constrained FBPINN predictions and reference solution in the microstructure region for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Comparison of the hard-constrained FBPINN predictions and reference solution in the microstructure region for the velocity ( [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]
Figure 18
Figure 18. Figure 18: FBPINN model predictions for the vertical velocity component for five di [PITH_FULL_IMAGE:figures/full_fig_p029_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 60 canonical work pages

  1. [1]

    B. Zhao, D. Sun, H. Wu, C. Qin, Q. Fei, Physics-informed neural networks for solving inverse problems in phase field models, Neural Networks 190 (2025) 107665

    work page 2025

  2. [2]

    J. M. Hanna, J. V . Aguado, S. Comas-Cardona, Y . Le Guennec, D. Borzacchiello, A self-supervised learning framework based on physics-informed and convolutional neural networks to identify local anisotropic perme- ability tensor from textiles 2D images for filling pattern prediction, Composites Part A: Applied Science and Manufacturing 179 (2024) 108019

    work page 2024

  3. [3]

    H. Moon, S. Lee, W. Demeke, B. Ryu, S. Ryu, Physics-informed neural operators for generalizable and label- free inference of temperature-dependent thermoelectric properties, npj Computational Materials 11 (1) (2025) 272

    work page 2025

  4. [4]

    Korolev, T

    D. Korolev, T. Schmidt, D. K. Natarajan, S. Cassola, D. May, M. Duhovic, M. Hintermüller, Hybrid machine learning based scale bridging framework for permeability prediction of fibrous structures, Composites Part A: Applied Science and Manufacturing 202 (2026) 109458

    work page 2026

  5. [5]

    J. Lee, M. Duhovic, D. May, T. Allen, P. Kelly, Physics-informed neural networks for real-time simulation of transverse Liquid Composite Moulding processes and permeability measurements, Composites Part A: Applied Science and Manufacturing 193 (2025) 108857

    work page 2025

  6. [6]

    Haghighat, S

    E. Haghighat, S. Abouali, R. Vaziri, Constitutive model characterization and discovery using physics-informed deep learning, Engineering Applications of Artificial Intelligence 120 (2023) 105828

    work page 2023

  7. [7]

    Hintermüller, D

    M. Hintermüller, D. Korolev, A hybrid physics-informed neural network based multiscale solver as a partial dif- ferential equation constrained optimization problem, ESAIM: Control, Optimisation and Calculus of Variations 32 (2026) 18. 33

    work page 2026

  8. [8]

    Shukla, Z

    K. Shukla, Z. Zou, C. H. Chan, A. Pandey, Z. Wang, G. E. Karniadakis, NeuroSEM: A hybrid framework for simulating multiphysics problems by coupling PINNs and spectral elements, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117498

    work page 2025

Show all 73 references
  1. [9]

    M. Asif, S. Jamshed, A. Dhiman, Heat transfer across an array of cylinders arranged in inline and staggered formation in a heat exchanger: Effect of nanoparticle volume fraction, nanoparticle diameter, and Richardson number, The European Physical Journal Plus 139 (2024) 601

    2024

  2. [10]

    Krishnamurthy, Y

    S. Krishnamurthy, Y . Peles, Gas-liquid two-phase flow across a bank of micropillars, Physics of Fluids 19 (2007) 043302

    2007

  3. [11]

    Horgue, F

    P. Horgue, F. Augier, P. Duru, M. Prat, M. Quintard, Experimental and numerical study of two-phase flows in arrays of cylinders, Chemical Engineering Science 102 (2013) 335–345

    2013

  4. [12]

    E. Wu, B. Wang, S. Zhang, Y . Su, X. Chen, Microscale underfill dynamics and void formation of high-density flip-chip packaging: Experiments and simulations, Physics of Fluids 36 (2024) 032117

    2024

  5. [13]

    S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: A review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738

    2021

  6. [14]

    S. Cai, Z. Wang, F. Fuest, Y . J. Jeon, C. Gray, G. E. Karniadakis, Flow over an espresso cup: inferring 3- D velocity and pressure fields from tomographic background oriented Schlieren via physics-informed neural networks, Journal of Fluid Mechanics 915 (2021) A102

    2021

  7. [15]

    Y . Zhu, W. Kong, J. Deng, X. Bian, Physics-informed neural networks for incompressible flows with moving boundaries, Physics of Fluids 36 (1) (2024)

    2024

  8. [16]

    Botarelli, M

    T. Botarelli, M. Fanfani, P. Nesi, L. Pinelli, Using Physics-Informed neural networks for solving Navier-Stokes equations in fluid dynamic complex scenarios, Engineering Applications of Artificial Intelligence 148 (2025) 110347

    2025

  9. [17]

    S. Wang, S. Sankaran, X. Fan, P. Stinis, P. Perdikaris, Simulating three-dimensional turbulence with physics- informed neural networks, arXiv preprint arXiv:2507.08972 (2025)

    2025

  10. [18]

    J. M. Hanna, J. V . Aguado, S. Comas-Cardona, R. Askri, D. Borzacchiello, Residual-based adaptivity for two- phase flow simulation in porous media using physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 396 (2022) 115100

    2022

  11. [19]

    J. D. Toscano, V . Oommen, A. J. Varghese, Z. Zou, N. Ahmadi Daryakenari, C. Wu, G. E. Karniadakis, From PINNs to PIKANs: Recent advances in physics-informed machine learning, Machine Learning for Computa- tional Science and Engineering 1 (1) (2025) 1–43

    2025

  12. [20]

    Bodaghi, G

    M. Bodaghi, G. Catalanotti, N. Correia, On the statistics of transverse permeability of randomly distributed fibers, Composite Structures 158 (2016) 323–332

    2016

  13. [21]

    Griebel, M

    M. Griebel, M. Klitz, Homogenization and numerical simulation of flow in geometries with textile microstruc- tures, Multiscale Modeling & Simulation 8 (4) (2010) 1439–1460

    2010

  14. [22]

    A. Yong, A. Endruweit, A. George, D. May, Y . Aksoy, M. Ali, T. Allen, P. Baral, C. Betteridge, C. Brauner, et al., Towards standardisation of the out-of-plane permeability measurement for reinforcement textiles, Composites Part A: Applied Science and Manufacturing 190 (2025) 108630

    2025

  15. [23]

    Annamalai, V

    P. Annamalai, V . Gangipamula, D. A. Ashebir, M. A. Sattar, S. V . Lomov, D. May, M. Bodaghi, M. Nikzad, On the extension of in-plane permeability calibration to out-of-plane measurements: Advancements in addi- tively manufactured textile-structured porous media for liquid com...

    2025

  16. [24]

    Syerko, T

    E. Syerko, T. Schmidt, D. May, C. Binetruy, S. Advani, S. Lomov, L. Silva, S. Abaimov, N. Aissa, I. Akha- tov, et al., Benchmark exercise on image-based permeability determination of engineering textiles: Microscale predictions, Composites Part A: Applied Science and Manufactu...

    2023

  17. [25]

    H. Jo, S. Bae, H. Hong, W. Kim, S. S. Kim, Prediction of transverse permeability in representative volume elements with closely arranged fibers through the application of Delaunay-triangulation and electrical-circuit analogy, Composite Structures 334 (2024) 117984

    2024

  18. [26]

    Caglar, G

    B. Caglar, G. Broggi, M. A. Ali, L. Orgéas, V . Michaud, Deep learning accelerated prediction of the permeability of fibrous microstructures, Composites Part A: Applied Science and Manufacturing 158 (2022) 106973

    2022

  19. [27]

    J. G. Jean, G. Broggi, B. Caglar, An image-based deep learning framework for flow field prediction in arbitrary- sized fibrous microstructures, Composites Part A: Applied Science and Manufacturing 200 (2026) 109337

    2026

  20. [28]

    Schmidt, D

    T. Schmidt, D. K. Natarajan, M. Duhovic, S. Cassola, M. Nuske, D. May, Numerical data generation for building machine learning models for permeability estimation of fibrous structures, Polymer Composites 46 (2025) S104– S120

    2025

  21. [29]

    S. Wang, S. Sankaran, H. Wang, P. Perdikaris, An expert’s guide to training physics-informed neural networks, arXiv preprint arXiv:2308.08468 (2023)

    2023 arXiv

  22. [30]

    Y . Zhu, N. Zabaras, P. Koutsourelakis, P. Perdikaris, Physics-constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data, Journal of Computational Physics 394 (2019) 56–81

    2019

  23. [31]

    F. M. Rohrhofer, S. Posch, C. Gößnitzer, B. C. Geiger, Data vs. Physics: The Apparent Pareto Front of Physics- Informed Neural Networks, IEEE Access 11 (2023) 86252–86261

    2023

  24. [32]

    S. Wang, Y . Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055–A3081

    2021

  25. [33]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y . Bengio, A. Courville, On the spectral bias of neural networks, in: International conference on machine learning, PMLR, 2019, pp. 5301–5310

    2019

  26. [34]

    S. Wang, H. Wang, P. Perdikaris, On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 384 (2021) 113938

    2021

  27. [35]

    Sukumar, A

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

    2022

  28. [36]

    S. Liu, H. Zhongkai, C. Ying, H. Su, J. Zhu, Z. Cheng, A unified hard-constraint framework for solving geomet- rically complex PDEs, NeurIPS 35 (2022) 20287–20299

    2022

  29. [37]

    Straub, P

    C. Straub, P. Brendel, V . Medvedev, A. Rosskopf, Hard-constraining Neumann boundary conditions in physics- informed neural networks via Fourier feature embeddings, arXiv preprint arXiv:2504.01093 (2025)

    2025 arXiv

  30. [38]

    J. Berg, K. Nyström, A unified deep artificial neural network approach to partial differential equations in complex geometries, Neurocomputing 317 (2018) 28–41

    2018

  31. [39]

    Hintermüller, J

    M. Hintermüller, J. Ning, Constrained neural parameterization for optimization in function spaces, arXiv preprint arXiv:2606.00855 (2026)

    2026 arXiv

  32. [40]

    J. Wang, Y . Mo, B. Izzuddin, C.-W. Kim, Exact dirichlet boundary physics-informed neural network EPINN for solid mechanics, Computer Methods in Applied Mechanics and Engineering 414 (2023) 116184

    2023

  33. [41]

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, S. G. Johnson, Physics-informed neural networks with hard constraints for inverse design, SIAM Journal on Scientific Computing 43 (6) (2021) B1105–B1132. 35

    2021

  34. [42]

    Y . Xie, H. Chi, Y . Wang, Y . Ma, Physics-specialized neural network with hard constraints for solving multi- material diffusion problems, Computer Methods in Applied Mechanics and Engineering 430 (2024) 117223

    2024

  35. [43]

    L. Sun, H. Gao, S. Pan, J.-X. Wang, Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data, Computer Methods in Applied Mechanics and Engineering 361 (2020) 112732

    2020

  36. [44]

    Klawonn, M

    A. Klawonn, M. Lanser, J. Weber, Machine learning and domain decomposition methods - a survey, Computa- tional Science and Engineering 1 (2024) 2

    2024

  37. [45]

    A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural networks (XPINNs): A generalized space- time domain decomposition based deep learning framework for nonlinear partial differential equations, Commu- nications in Computational Physics 28 (5) (2020) 2002–2041

    2020

  38. [46]

    Moseley, A

    B. Moseley, A. Markham, T. Nissen-Meyer, Finite basis physics-informed neural networks (FBPINNs): a scal- able domain decomposition approach for solving differential equations, Advances in Computational Mathemat- ics 49 (4) (2023) 62

    2023

  39. [47]

    Z. Hu, A. D. Jagtap, G. E. Karniadakis, K. Kawaguchi, Augmented physics-informed neural networks (AP- INNs): A gating network-based soft domain decomposition methodology, Engineering Applications of Artificial Intelligence 126 (2023) 107183

    2023

  40. [48]

    J. K. Hunter, B. Nachtergaele, Applied analysis, World Scientific, 2001

    2001

  41. [49]

    Zeinhofer, R

    M. Zeinhofer, R. Masri, K.-A. Mardal, A unified framework for the error analysis of physics-informed neural networks, IMA Journal of Numerical Analysis 45 (5) (2025) 2988–3025

    2025

  42. [50]

    Mishra, R

    S. Mishra, R. Molinaro, Estimates on the generalization error of physics-informed neural networks for approxi- mating PDEs, IMA Journal of Numerical Analysis 43 (1) (2023) 1–43

    2023

  43. [51]

    Y . Shin, Z. Zhang, G. E. Karniadakis, Error estimates of residual minimization using neural networks for linear PDEs, Journal of Machine Learning for Modeling and Computing 4 (4) (2023) 73–101

    2023

  44. [52]

    Wolf, Homogenization of the Stokes system in a non-periodically perforated domain, Multiscale Modeling & Simulation 20 (1) (2022) 72–106

    S. Wolf, Homogenization of the Stokes system in a non-periodically perforated domain, Multiscale Modeling & Simulation 20 (1) (2022) 72–106

    2022

  45. [53]

    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 Computa- tional physics 378 (2019) 686–707

    2019

  46. [54]

    I. E. Lagaris, A. Likas, D. I. Fotiadis, Artificial neural networks for solving ordinary and partial differential equations, IEEE transactions on neural networks 9 (5) (1998) 987–1000

    1998

  47. [55]

    De Ryck, A

    T. De Ryck, A. D. Jagtap, S. Mishra, Error estimates for physics-informed neural networks approximating the Navier–Stokes equations, IMA Journal of Numerical Analysis 44 (1) (2024) 83–119

    2024

  48. [56]

    De Ryck, S

    T. De Ryck, S. Mishra, Numerical analysis of physics-informed neural networks and related models in physics- informed machine learning, Acta Numerica 33 (2024) 633–713

    2024

  49. [57]

    Müller, M

    J. Müller, M. Zeinhofer, Notes on exact boundary values in residual minimisation, in: Mathematical and Scien- tific Machine Learning, PMLR, 2022, pp. 231–240

    2022

  50. [58]

    Johnson, Numerical solution of partial differential equations by the finite element method, Courier Corpora- tion, 2009

    C. Johnson, Numerical solution of partial differential equations by the finite element method, Courier Corpora- tion, 2009

    2009

  51. [59]

    Gühring, M

    I. Gühring, M. Raslan, Approximation rates for neural networks with encodable weights in smoothness spaces, Neural Networks 134 (2021) 107–130. 36

    2021

  52. [60]

    Bischof, M

    R. Bischof, M. A. Kraus, Multi-objective loss balancing for physics-informed deep learning, Computer Methods in Applied Mechanics and Engineering 439 (2025) 117914

    2025

  53. [61]

    Wright, J

    S. Wright, J. Nocedal, Numerical optimization, Springer Science 35 (67-68) (1999) 7

    1999

  54. [62]

    Rathore, W

    P. Rathore, W. Lei, Z. Frangella, L. Lu, M. Udell, Challenges in training PINNs: a loss landscape perspective, in: Proceedings of the 41st International Conference on Machine Learning, ICML, Vienna, Austria, 2024

    2024

  55. [63]

    T. D. Ryck, F. Bonnet, S. Mishra, E. de Bézenac, An operator preconditioning perspective on training in physics- informed machine learning, in: ICLR, 2024

    2024

  56. [64]

    Q. Liu, M. Chu, N. Thuerey, Config: Towards conflict-free training of physics informed neural networks, arXiv preprint arXiv:2408.11104 (2024)

    2024 arXiv

  57. [65]

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, C. Finn, Gradient surgery for multi-task learning, Advances in neural information processing systems 33 (2020) 5824–5836

    2020

  58. [66]

    S. Wang, A. K. Bhartari, B. Li, P. Perdikaris, Gradient alignment in physics-informed neural networks: A second-order optimization perspective, arXiv preprint arXiv:2502.00604 (2025)

    2025

  59. [67]

    C. Xu, B. T. Cao, Y . Yuan, G. Meschke, Transfer learning based physics-informed neural networks for solving inverse problems in engineering structures under different loading scenarios, Computer Methods in Applied Mechanics and Engineering 405 (2023) 115852

    2023

  60. [68]

    D. P. Kingma, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  61. [69]

    G. B. Folland, A. Sitaram, The uncertainty principle: a mathematical survey, Journal of Fourier analysis and applications 3 (3) (1997) 207–238

    1997

  62. [70]

    L. Lu, X. Meng, Z. Mao, G. E. Karniadakis, DeepXDE: A Deep Learning Library for solving Differential Equations, SIAM Review 63 (1) (2021) 208–228

    2021

  63. [71]

    C. Wu, M. Zhu, Q. Tan, Y . Kartha, L. Lu, A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 403 (2023) 115671

    2023

  64. [72]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, et al., JAX: composable transformations of Python+NumPy programs (2018)

    2018

  65. [73]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, R. Ng, Fourier features let networks learn high frequency functions in low dimensional domains, Advances in neural information processing systems 33 (2020) 7537–7547. 37

    2020

Pith tools

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