Pith. sign in

REVIEW 2 major objections 5 minor 52 references

LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Integrated Cauchy ridge atoms match the value–gradient–curvature structure of thin convection layers, giving thickness-uniform exponential approximation and more accurate PINNs with far fewer parameters.

desk verdict Solid representation-theory paper: integrated Cauchy ridges give derivative-stable, thickness-uniform layer approximation, and the numerics back the story without overclaiming optimizer independence. read the letter →

arxiv 2607.03682 v2 pith:GSVADMLX submitted 2026-07-04 math.AP cs.LG

Zihao Guo , Xin Li , Zhihong Xia This is my paper · ORCID
classification math.APcs.LG MSC 65N3535B2565D1568T07
keywords Physics-informedneuralnetworksConvection-dominatedconvection–diffusionequationsIntegratedCauchyactivationXNetLayer-resolvingrepresentationSingularlyperturbedproblemsRidgeneurons
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

Convection-dominated diffusion problems form thin layers whose values stay order one while gradients and curvatures scale like inverse powers of the layer thickness. Standard physics-informed networks do not encode that coupled structure, so accuracy collapses as diffusion shrinks. This paper builds ridge neurons from integrated Cauchy activations: each atom is a transition at the solution level, and its first and second derivatives are localized Cauchy profiles whose physical width is d over the weight norm. For analytic stretched profiles, one set of atoms approximates the profile and its first two derivatives at a shared exponential rate; after rescaling, the constants stay independent of thickness for both O(ε) exponential and O(√ε) characteristic layers. On interior, boundary, outflow, corner, and curved-layer benchmarks the resulting LRX-PINN outperforms larger Kolmogorov–Arnold and Fourier-feature networks while using under 30 percent of their parameters, and the same basis improves published variational PINN baselines when inserted without changing losses or stabilizations.

What carries the argument

The geometric integrated Cauchy atom Ψ = A arctan((n·x−s)/ρ) + (B/2) log((n·x−s)²+ρ²), with physical width ρ=d/‖w‖. Its derivatives recover localized Cauchy kernels of amplitudes ρ⁻¹ and ρ⁻²; Theorem 1 shows one M-term sum approximates U, U′, U″ at rate σ⁻ᴹ, which Theorems 2–3 lift to thickness-uniform physical norms for both O(ε) and O(√ε) layers.

What would settle it

For a known analytic layer of thickness δ, compute best M-term integrated Cauchy error in the scaled X_δ norm; if that error grows as δ shrinks at fixed M, or fails to decay exponentially in M with δ-independent constants, the thickness-uniform claim is false.

Watch

Extended reading notes

Core claim

An integrated Cauchy ridge trial space matches the value–gradient–curvature hierarchy of convection–diffusion layers: one collection of atoms simultaneously approximates an analytic stretched profile and its first two derivatives at a common exponential rate, and after physical rescaling the approximation constants do not deteriorate as layer thickness tends to zero. This yields thickness-uniform best-approximation estimates for exponential and characteristic layers and a layer-scaled residual bound for the singularly perturbed operator, realized as LRX-PINN in strong form and as an embeddable trial space in stabilized variational PINNs.

Load-bearing premise

The exponential rates assume each stretched layer profile extends holomorphically into a fixed complex neighborhood whose size does not shrink with the physical layer thickness.

Editorial extensions

If this is right

  • Best-approximation complexity for a fixed number of separated analytic layers grows only logarithmically with target accuracy and is independent of inverse minimum layer thickness.
  • Strong-form PINNs and weak or hp-VPINN formulations can share the same layer-adapted trial space; the strong residual uses the full value–gradient–curvature hierarchy while the variational form needs only value and first derivative.
  • Matching neuron physical width to O(ε) or O(√ε) by layer type removes the need to inflate network size solely because layers become thinner.
  • On the reported benchmarks accuracy improves while trainable parameters drop below 30 percent of competing PINN architectures; embedding the basis into existing stabilized hp-VPINNs further lowers published error floors without retuning losses.

Reading between the lines

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

  • The explicit physical width ρ could be monitored during training as an automatic layer detector, linking neural representation to classical layer-adapted mesh indicators.
  • Closed-form residual expressions in the integrated Cauchy basis and its derivatives may reduce automatic-differentiation cost in higher dimensions relative to deep generic networks.
  • Extending beyond linear steady problems first requires checking whether nonlinear layer profiles retain a sufficient complex neighborhood for the Cauchy-quadrature argument.
  • Different ρ groups for coexisting exponential and characteristic layers suggest a multi-scale dictionary that could be sparsified by group selection rather than deeper stacking.
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

2 major / 5 minor

Summary. The paper constructs an integrated Cauchy ridge trial space for convection-dominated convection–diffusion equations. Each atom is transition-type at the solution level, while its first and second derivatives recover localized Cauchy-type profiles; the parameters encode normal direction, location, and physical width ρ=d/∥w∥. For analytic stretched layer profiles, Theorem 1 proves that one M-term integrated Cauchy sum approximates U, U′, and U′′ at a common exponential rate via Cauchy-integral trapezoidal quadrature; Theorems 2–3 and Corollaries 1–2 then give thickness-uniform scaled L∞/H² estimates for both O(ε) exponential and O(√ε) characteristic layers, with residual consequences for strong and variational forms. The same basis is realized as LRX-PINN and embedded into existing hp-VPINN frameworks without changing their losses or stabilization. Numerics on interior, boundary, mixed, outflow, corner, and curved layers report higher accuracy than PIKAN and Fourier PINNs with far fewer parameters, and further gains over published hp-VPINN baselines.

Significance. If the claims hold, the work supplies a concrete, analyzable neural trial space matched to the value–gradient–curvature hierarchy of singularly perturbed layers, rather than only a new residual or training heuristic. The derivative-stable exponential approximation (Thm. 1) and thickness-uniform physical-space bounds (Thms. 2–3) are the main theoretical contribution; the numerical design (matched training, original vs integrated activation, embedding into unchanged hp-VPINN losses, seed robustness in Table 19, closed-form residual acceleration) is a genuine strength and separates representation effects from loss design. The paper is careful to separate best-approximation complexity from optimizer/conditioning costs (Remark 3, Thm. 4 discussion). This is a solid representation-oriented contribution for PINNs on convection-dominated problems.

major comments (2)
  1. The exponential rates in Lemma 1 / Theorem 1 rest on holomorphic extension of U (and of the contour parametrization) into a fixed complex neighborhood with constants independent of physical thickness δ. The manufactured benchmarks satisfy this; many practical layers (e.g., non-analytic or only C^k profiles, or layers with ε-dependent analyticity strips) may not. The paper already flags the representation–optimizer gap (Remark 3). A short discussion or one non-analytic / reduced-regularity numerical example would clarify the practical scope of the thickness-uniform residual claims without changing the theorems.
  2. Theorem 4 and Corollaries 1–2 cover coexisting O(ε) and O(√ε) layers via different atom widths, but the numerical suite does not include an explicit multi-scale manufactured solution with both exponential and characteristic layers simultaneously (Tasks 2–4 use a single scale; Tasks 5–6 are extreme but single-scale or curved). Adding one such experiment, or clearly stating that multi-scale coexistence is left to theory, would better support the multi-scale complexity claim.
minor comments (5)
  1. Section cross-references appear as Section?? in the introduction and organization paragraph; fix numbering.
  2. Abstract and title use LRX-PINN / XNet branding; the body title is “Layer-Resolving Rational Trial-Space Method…”. Align titles and keywords for consistency.
  3. Table 1 is a useful structural summary; a brief pointer from the theory section to the corresponding tasks would help readers map analysis to experiments.
  4. Initial scale d is task- and ε-dependent (Table 18, Fig. 16). State the selection protocol more explicitly in §3.3 so that the free-parameter list is transparent.
  5. Minor typesetting: missing spaces in compound words (e.g., “Physics-informedneuralnetworks”, “convection–diffusionproblems”) and occasional broken math spacing in the introduction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: approximation rates are classical complex-analysis best-approximation results; numerics compare against external baselines under fixed losses.

full rationale

The load-bearing theoretical chain is Lemma 1 (Cauchy-integral trapezoidal quadrature for holomorphic F) applied to U′, integration to obtain GM, and Cauchy estimates for U″ (Theorem 1), then rescaling to physical thickness δ (Theorems 2–3, Corollaries 1–2). These are standard complex-analysis arguments under stated holomorphic-extension hypotheses; they do not define the approximant from the residual they later bound, nor do they fit parameters to residual data and relabel the fit as a prediction. Self-citations to XNet/CompleX PINN ([28,29,41]) motivate the Cauchy family but are not used as uniqueness theorems that force the layer-scaled rates; the integrated-Cauchy analysis is re-derived in the paper. Numerical claims compare LRX-PINN to external PIKAN, Fourier-feature PINN, and published hp-VPINN numbers, and for Tasks 5–6 the original loss functionals and stabilization are held fixed so only the trial space changes. Empirical choice of initial d is ordinary hyperparameter selection, not a fitted-input-called-prediction. Remark 3 and the discussion of Theorem 4 explicitly separate best-approximation complexity from optimizer/conditioning costs. No step reduces a claimed prediction to its inputs by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The load-bearing math is standard complex approximation plus singular-perturbation layer scaling; the paper’s own additions are the integrated Cauchy atom geometry and the thickness-uniform transfer. Free parameters are architectural/training choices (neuron count, initial d, loss weights, iteration budgets) that affect realized accuracy but not the best-approximation theorems. No new physical particles or forces are postulated—only a neural basis.

free parameters (5)
  • N_Cauchy (hidden neurons)
    Fixed at 120 for Tasks 1–4; directly sets representation capacity and the 841-parameter count used in all efficiency claims.
  • initial activation scale d (per task/ε)
    Table 18 reports best-performing initial d decreasing with ε; chosen by ablation after 5000 Adam steps, so numerical wins partly depend on this hand-tuned scale.
  • loss weights λ_f, λ_b
    Fixed at 1 and 40 for Tasks 1–4 without adaptive balancing; boundary emphasis is a free training choice.
  • Adam/L-BFGS iteration budgets
    8000+12000 for Tasks 1–4; 50k–100k Adam for Tasks 5–6 following prior papers—optimizer effort is prescribed, not derived.
  • Fourier scale σ and KAN/Fourier widths
    Baseline hyperparameters (σ by ε; widths 20/32, etc.) are configuration choices that frame the parameter-efficiency comparison.
assumptions (4)
  • domain assumption Stretched layer profiles U are holomorphic in a fixed complex neighborhood of a compact real interval with conjugate symmetry, enabling trapezoidal Cauchy quadrature exponential rates.
    Hypothesis of Lemma 1 and Theorem 1; without a δ-independent analyticity strip the exponential σ^{-M} rate is not guaranteed.
  • domain assumption Local layer structure is essentially one-dimensional in a normal coordinate (planar patch or tubular neighborhood for curved layers).
    Section 2.1 and Remark 4; corner interactions and strong curvature corrections are treated as lower-order or multi-atom superpositions.
  • standard math Cauchy integral formula and trapezoidal-rule error bounds for 2π-periodic holomorphic functions on a strip.
    Lemma 1 proof; classical complex analysis used as black box.
  • ad hoc to paper Best-approximation complexity is separate from nonconvex optimization, conditioning, and quadrature cost as ε→0.
    Explicitly stated after Theorem 4 and Remark 3; numerical claims still rely on Adam/L-BFGS finding good atoms.
invented entities (1)
  • Integrated Cauchy ridge atom / VIC_M trial space (LRX-PINN basis) independent evidence
    purpose: Provide a single parametric family that is transition-type in value and Cauchy-localized in first/second derivatives with explicit physical width ρ=d/‖w‖.
    Mathematical construction built from known Cauchy kernels; independent mathematical content is the approximation theory, not external physical detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems." pith.science (2026). https://pith.science/paper/GSVADMLX

@misc{pith2026260703682,
  author       = {Pith},
  title        = {Pith review of: LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSVADMLX}},
  note         = {Machine review of arXiv:2607.03682}
}
abstract

Convection-dominated convection-diffusion problems often develop thin layers, where the solution has sharp transition profiles and its derivatives are highly localized. This creates a structural mismatch for standard physics-informed neural networks (PINNs), whose trial spaces are not designed to match the value--derivative structure of such layers. We propose a Layer-Resolving XNet Physics-Informed Neural Network (LRX-PINN) based on integrated Cauchy activations. The proposed basis is transition-type at the solution level, while its derivative recovers a localized Cauchy kernel. We show that this structure matches the scaling of convection-dominated layers, inherits the Cauchy approximation mechanism at the derivative-profile level, and identifies \(d/\|w\|\) as the effective physical width of a ridge neuron. For analytic layer profiles, this yields derivative-stable exponential approximation in the stretched coordinate and a layer-scaled estimate for the strong residual of the singularly perturbed operator. Numerical experiments on several convection-dominated benchmarks show that LRX-PINN achieves higher accuracy than PIKAN and Fourier-feature PINNs while using less than \(30\%\) of their trainable parameters. On more challenging benchmarks, embedding the proposed representation into hp-VPINN-based frameworks further improves the best results obtained by existing hp-VPINN-based baselines without changing their original loss functionals or stabilization strategies. These results show that neural representations aligned with layer structure provide a compact and effective approach for convection-dominated problems.

Figures

Figures reproduced from arXiv: 2607.03682 by the authors.

Figure 1
Figure 1. Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 1
Figure 1. Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figures from the paper (29 more)
Figure 2
Figure 2. Figure 2: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p021_2.png]
Figure 3
Figure 3. Figure 3: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p015_3.png]
Figure 3
Figure 3. Figure 3: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p022_3.png]
Figure 4
Figure 4. Figure 4: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 4
Figure 4. Figure 4: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 5
Figure 5. Figure 5: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 6
Figure 6. Figure 6: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 7
Figure 7. Figure 7: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 8
Figure 8. Figure 8: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 9
Figure 9. Figure 9: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 10
Figure 10. Figure 10: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 11
Figure 11. Figure 11: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 12
Figure 12. Figure 12: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 12
Figure 12. Figure 12: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]
Figure 13
Figure 13. Figure 13: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 13
Figure 13. Figure 13: Exact solution (left), LRX-PINN prediction (middle), and absolute error (right) for [PITH_FULL_IMAGE:figures/full_fig_p028_13.png]
Figure 14
Figure 14. Figure 14: Exact solution (left), LRX-enhanced hp-VPINN prediction with the integrated [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 14
Figure 14. Figure 14: Exact solution (left), LRX-enhanced hp-VPINN prediction with the integrated [PITH_FULL_IMAGE:figures/full_fig_p029_14.png]
Figure 15
Figure 15. Figure 15: Exact solution (left), LRX-enhanced hp-VPINN prediction with the integrated [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]
Figure 15
Figure 15. Figure 15: Exact solution (left), LRX-enhanced hp-VPINN prediction with the integrated [PITH_FULL_IMAGE:figures/full_fig_p030_15.png]
Figure 16
Figure 16. Figure 16: Optimal initial scale parameter d obtained from the ablation study under different parameters ε. We further evaluate robustness with respect to random initialization by repeating each ex￾periment using six different random seeds 0-5 [PITH_FULL_IMAGE:figures/full_fig_…
Figure 16
Figure 16. Figure 16: Optimal initial scale parameter d obtained from the ablation study under different parameters ε. We further evaluate robustness with respect to random initialization by repeating each ex￾periment using six different random seeds 0-5 [PITH_FULL_IMAGE:figures/full_fig_…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 1 linked inside Pith

  1. [1]

    Anandh, D

    T. Anandh, D. Ghose, H. Jain, P. Sunkad, S. Ganesan, V. John, Improving hp-Variational Physics-Informed Neural Networks for Steady-State Convection-Dominated Problems, Computer Methods in Applied Mechanics and Engineering, 438 (2025), 117797

    2025

  2. [2]

    Augustin, A

    M. Augustin, A. Caiazzo, A. Fiebach, J. Fuhrmann, V. John, A. Linke, R. Umla, An assess- ment of discretizations for convection-dominated convection-diffusion equations,Computer Methods in Applied Mechanics and Engineering, 200 (47/48) (2011), 3395–3409

    2011

  3. [3]

    G. R. Barrenechea, V. John, P. Knobloch, Finite element methods respecting the discrete maximum principle for convection–diffusion equations,SIAM Review, 66 (2024), 3–88

    2024

  4. [4]

    Bear,Dynamics of Fluids in Porous Media, Elsevier, 1972

    J. Bear,Dynamics of Fluids in Porous Media, Elsevier, 1972

    1972

  5. [5]

    R. B. Bird, W. E. Stewart, E. N. Lightfoot,Transport Phenomena, John Wiley & Sons, 2002

    2002

  6. [6]

    Bos and N

    L. Bos and N. Levenberg, Bernstein–Walsh theory associated to convex bodies and applica- tions to multivariate approximation theory,Computational Methods and Function Theory, 18 (2018), 361–388

    2018

  7. [7]

    A.N.Brooks, T.J.R.Hughes, Streamlineupwind/Petrov–Galerkinformulationsforconvec- tion dominated flows with particular emphasis on the incompressible Navier–Stokes equa- tions,Computer Methods in Applied Mechanics and Engineering, 32 (1982), 199–259

    1982

  8. [8]

    S. Cai, Z. Wang, S. Wang, P. Perdikaris, G. E. Karniadakis, Physics-informed neural net- works for heat transfer problems,Journal of Heat Transfer, 143 (2021), 060801

    2021

Show all 52 references
  1. [9]

    R. Codina, Stabilized finite element approximation of transient incompressible flows using orthogonalsubscales,Computer Methods in Applied Mechanics and Engineering, 191(39–40) (2002), 4295–4321

    2002

  2. [10]

    De Paoli, Convective mixing in porous media: A review of Darcy, pore-scale and Hele- Shaw studies,European Physical Journal E, 46 (2023), 1–41

    M. De Paoli, Convective mixing in porous media: A review of Darcy, pore-scale and Hele- Shaw studies,European Physical Journal E, 46 (2023), 1–41

    2023

  3. [11]

    R. A. DeVore and G. G. Lorentz,Constructive Approximation, Springer, Berlin, 1993

    1993

  4. [12]

    L. C. Evans,Partial Differential Equations, American Mathematical Society, 2010

    2010

  5. [13]

    J. H. Ferziger, M. Perić,Computational Methods for Fluid Dynamics, Springer, 2002. 34

    2002

  6. [14]

    Frerichs-Mihov, L

    D. Frerichs-Mihov, L. Henning, V. John, On Loss Functionals for Physics-Informed Neural Networks for Steady-State Convection-Dominated Convection-Diffusion Problems,Commu- nications on Applied Mathematics and Computation, 8 (2026), 287–308

    2026

  7. [15]

    Giera, T

    S. Giera, T. Iliescu, V. John, D. Wells, SUPG reduced order models for convection- dominated convection–diffusion–reaction equations,Computer Methods in Applied Mechan- ics and Engineering, 289 (2015), 454–474

    2015

  8. [16]

    Haghighat, R

    E. Haghighat, R. Juanes, SciANN: A Keras wrapper for scientific computations and physics- informed deep learning using artificial neural networks,Computer Methods in Applied Me- chanics and Engineering, 373 (2021), 113552

    2021

  9. [17]

    T. J. R. Hughes, L. P. Franca, and G. M. Hulbert, A new finite element formulation for computational fluid dynamics: VIII. The Galerkin/least-squares method for advective- diffusive equations,Computer Methods in Applied Mechanics and Engineering, 73(2) (1989), 173–189

    1989

  10. [18]

    T. J. R. Hughes, Multiscale phenomena: Green’s functions, the Dirichlet-to-Neumann for- mulation, subgrid scale models, bubbles and the origins of stabilized methods,Computer Methods in Applied Mechanics and Engineering, 127(1–4) (1995), 387–401

    1995

  11. [19]

    Jackson, On approximation by trigonometric sums and polynomials,Transactions of the American Mathematical Society, 13 (1912), 491–515

    D. Jackson, On approximation by trigonometric sums and polynomials,Transactions of the American Mathematical Society, 13 (1912), 491–515

    1912

  12. [20]

    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,Communications in Computational Physics, 28 (2020), 2002– 2041

    2020

  13. [21]

    X. Jin, S. Cai, H. Li, G. E. Karniadakis, NSFnets (Navier–Stokes flow nets): Physics- informed neural networks for the incompressible Navier–Stokes equations,Journal of Com- putational Physics, 426 (2021), 109951

    2021

  14. [22]

    John, J.M

    V. John, J.M. Maubach, L. Tobiska, Nonconforming streamline-diffusion-finite-element- methods for convection-diffusion problems,Numerische Mathematik, 78 (2) (1997), 165– 188

    1997

  15. [23]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics- informed machine learning,Nature Reviews Physics, 3 (2021), 422–440

    2021

  16. [24]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, G. E. Karniadakis, hp-VPINNs: Variational physics-informed neu- ral networks with domain decomposition, Computer Methods in Applied Mechanics and Engineering, 374 (2021), 113547

    2021

  17. [25]

    Klimek,Pluripotential Theory, Oxford University Press, Oxford, 1991

    M. Klimek,Pluripotential Theory, Oxford University Press, Oxford, 1991

    1991

  18. [26]

    Kraus, Bernstein–Walsh type theorems for real analytic functions in several variables, Constructive Approximation, 33 (2011), 191–217

    C. Kraus, Bernstein–Walsh type theorems for real analytic functions in several variables, Constructive Approximation, 33 (2011), 191–217

    2011

  19. [27]

    A. S. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. Mahoney, Characterizing possible failure modes in physics-informed neural networks, Advances in Neural Information Pro- cessing Systems, 34 (2021), 26548–26560

    2021

  20. [28]

    X. Li, Z. Xia, H. Zhang, Cauchy activation function and XNet,Neural Networks, 190 (2025), 107730. 35

    2025

  21. [29]

    X. Li, X. Zheng, Z. Xia, From Kolmogorov to Cauchy: Shallow XNet Surpasses KANs, Advances in Neural Information Processing Systems, 38 (2025), 129075–129100

    2025

  22. [30]

    Z. Li, D. Z. Huang, B. Liu, A. Anandkumar, Fourier neural operator with learned deforma- tions for PDEs on general geometries,Journal of Machine Learning Research, 24 (2023), 1–26

    2023

  23. [31]

    Linß,Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Springer, 2010

    T. Linß,Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Springer, 2010

    2010

  24. [32]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljačić, T. Y. Hou, M. Tegmark, KAN: Kolmogorov–Arnold networks,International Conference on Learning Representa- tions, 2025

    2025

  25. [33]

    Matthaiou, V

    M. Matthaiou, V. John, M. Zainelabdeen, Bound-preserving physics-informed neural net- works for steady-state convection-diffusion-reaction problems,Computers & Mathematics with Applications, 199 (2025), 167–183

    2025

  26. [34]

    K. W. Morton,Numerical Solution of Convection–Diffusion Problems, Chapman & Hall, 1996

    1996

  27. [35]

    D. J. Newman, Rational approximation to|x|,Michigan Mathematical Journal, 11(1) (1964), 11–14

    1964

  28. [36]

    Patra, S

    S. Patra, S. Panda, B. K. Parida, M. Arya, K. Jacobs, D. I. Bondar, A. Sen, Physics- informed Kolmogorov–Arnold neural networks for dynamical analysis via Efficient-KAN and WAV-KAN,Journal of Machine Learning Research, 26 (2025), 1–39

    2025

  29. [37]

    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 Computational Physics, 378 (2019), 686–707

    2019

  30. [38]

    H.-G. Roos, M. Stynes, L. Tobiska,Robust Numerical Methods for Singularly Perturbed Differential Equations, Springer, 2008

    2008

  31. [39]

    Runje, S

    D. Runje, S. M. Shankaranarayana, Constrained monotonic neural networks,Proceedings of the 40th International Conference on Machine Learning (ICML), 202 (2023), 29338–29353

    2023

  32. [40]

    Sartor, A

    D. Sartor, A. Sinigaglia, G. A. Susto, Advancing constrained monotonic neural networks: Achieving universal approximation beyond bounded activations,Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025

    2025

  33. [41]

    C. Si, M. Yan, X. Li, Z. Xia, CompleX physics-informed neural networks,Journal of Com- putational Physics, 2026, 114713

    2026

  34. [42]

    J. Siciak, On some extremal functions and their applications in the theory of analytic functions of several complex variables,Transactions of the American Mathematical Society, 105 (1962), 322–357

    1962

  35. [43]

    Sikora, P

    M. Sikora, P. Krukowski, A. Paszynska, M. Paszynski, Physics informed neural networks with strong and weak residuals for advection-dominated diffusion problems,arXiv preprint arXiv:2307.07647, (2023)

    2023 arXiv

  36. [44]

    J.Sirignano, K.Spiliopoulos, DGM:Adeeplearningalgorithmforsolvingpartialdifferential equations,Journal of Computational Physics, 375 (2018), 1339–1364

    2018

  37. [45]

    C. Song, Y. Wang, Simulating seismic multifrequency wavefields with the Fourier feature physics-informed neural network,Geophysical Journal International, 232 (2023), 1503–1514. 36

    2023

  38. [46]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. 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

    2020

  39. [47]

    J. D. Toscano, V. Oommen, A. J. Varghese, Z. Zou, N. A. Daryakenari, C. Wu, G. E. Kar- niadakis, From PINNs to PIKANs: Recent advances in physics-informed machine learning, Machine Learning for Computational Science and Engineering, 1 (2025), 15

    2025

  40. [48]

    J. L. Walsh,Interpolation and Approximation by Rational Functions in the Complex Do- main, 5th ed., American Mathematical Society, Providence, RI, 1969

    1969

  41. [49]

    S. Wang, S. Sankaran, H. Wang, P. Perdikaris, Respecting causality for training physics- informed neural networks,Computer Methods in Applied Mechanics and Engineering, 421 (2024), 116813

    2024

  42. [50]

    S. Wang, X. Yu, P. Perdikaris, When and why PINNs fail to train: A neural tangent kernel perspective,Journal of Computational Physics, 449 (2022), 110768

    2022

  43. [51]

    Y. Wang, C. Xu, M. Yang, J. Zhang, Less emphasis on hard regions: Curriculum learning of PINNs for singularly perturbed convection–diffusion–reaction problems,East Asian Journal on Applied Mathematics, 14 (2024), 104–123

    2024

  44. [52]

    Y. Wang, J. Sun, J. Bai, C. Anitescu, M. S. Eshaghi, X. Zhuang, T. Rabczuk, Y. Liu, Kolmogorov–Arnold-informed neural network: A physics-informed deep learning framework for solving forward and inverse problems,Computer Methods in Applied Mechanics and Engineering,423 (2025), ...

    2025

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.