pith. sign in

arxiv: 2507.21437 · v2 · submitted 2025-07-29 · 💻 cs.LG

PVD-ONet: A Multi-scale Neural Operator Method for Singularly Perturbed Boundary Layer Problems

Tiantian Sun , Jian Zu This is my paper

Pith reviewed 2026-05-19 03:06 UTC · model grok-4.3

classification 💻 cs.LG
keywords physics-informed neural networkssingularly perturbed problemsboundary layersoperator learningPrandtl matchingVan Dyke principlemulti-scale methods
0
0 comments X

The pith

Prandtl-Van Dyke matching conditions embedded in neural architectures solve singularly perturbed boundary layer problems where standard physics-informed networks fail.

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

The paper develops PVD-Net and its operator-learning extension PVD-ONet to handle singularly perturbed PDEs that produce thin boundary layers. It builds two versions of the network: a two-network setup that uses Prandtl's matching condition for stability, and a five-network setup that applies Van Dyke's matching principle for higher accuracy. Both versions enforce the governing equations through the loss without requiring solution data. The operator version assembles multiple DeepONet modules so that a single trained model maps initial conditions to solutions across families of related problems. Experiments on constant-coefficient, variable-coefficient, and internal-layer cases show consistent gains over existing baselines, and the same framework can recover the layer-thickness scaling exponent from sparse observations.

Core claim

PVD-Net uses a two-network architecture together with Prandtl's matching condition to produce stable solutions for stability-focused tasks and a five-network architecture together with Van Dyke's matching principle to resolve fine-scale layer structures for accuracy-focused tasks; PVD-ONet generalizes the same idea by composing multiple DeepONet modules that learn the solution operator directly from the governing equations, enabling instant evaluation on new initial conditions without retraining.

What carries the argument

Multi-network neural architectures that enforce Prandtl or Van Dyke matching conditions between inner boundary-layer and outer solutions as soft constraints in the loss.

If this is right

  • The methods produce more accurate solutions than existing physics-informed baselines on second-order equations with both constant and variable coefficients.
  • The same architectures handle internal-layer problems without special tuning.
  • PVD-ONet delivers instant predictions for entire families of boundary-layer problems once trained.
  • The framework recovers the unknown scaling exponent that controls boundary-layer thickness from sparse data alone.

Where Pith is reading between the lines

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

  • The same matching-based decomposition could be applied to other multi-scale problems such as reaction-diffusion systems or high-Reynolds-number flows.
  • Extending the approach to three-dimensional or time-dependent singularly perturbed equations would test its generality beyond the one-dimensional cases shown.
  • Because the operator version learns families of solutions, it may serve as a fast surrogate inside optimization loops that vary the perturbation parameter.

Load-bearing premise

The classical Prandtl and Van Dyke matching rules remain accurate and stable when they are added as soft constraints inside the neural-network loss for the singularly perturbed problems considered.

What would settle it

Numerical failure to converge or large pointwise errors on the standard convection-diffusion test problem with small diffusion coefficient and known exact solution would falsify the claim.

Figures

Figures reproduced from arXiv: 2507.21437 by Jian Zu, Tiantian Sun.

Figure 1
Figure 1. Figure 1: Schematic diagram of the boundary layer scale transformation. The diagram illustrates the evolution of the boundary [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of predicted results in the boundary layer region between our method and the baseline (BL-PINNs), using [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The network architecture of PVD-Net. The PVD-Net consists of two distinct configurations: a leading-order ver [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The architecture of Physics-informed DeepONet. The network consists of a branch network that encodes the input [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The architecture of PVD-ONet. Similar to PVD-Net, PVD-ONet also consists of two approximation versions. PVD [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (Left): Predicted results of Leading-order PVD-Net. (Right): Predicted results of High-order PVD-Net. As can be seen from the above panels, our proposed PVD-Net is able to efficiently learn the solutions of the boundary layer problem, and the differences between the analytic and real solutions are almost indistinguishable, while these solutions exhibit smooth transitions. It is particularly notable that th… view at source ↗
Figure 7
Figure 7. Figure 7: Assuming the location of the boundary layer is known, we separately employ DeepONets to learn the inner and [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (Left): Predicted results of Leading-order PVD-ONet. (Right): Predicted results of High-order PVD-ONet. We aim to learn the operator G : (a, b) 7→ u. As shown, the proposed PVD-ONet is capable of accurately learning the mapping from parameters to solutions in boundary layer problems. The predicted solutions closely match the ground truth, exhibiting smooth transitions. This proves that our proposed framewo… view at source ↗
Figure 9
Figure 9. Figure 9: Prediction results of the PVD-Net for the second-order differential equation with variable coefficients. [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: PVD-ONet for the second-order differential equation with variable coefficients. [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗
read the original abstract

Physics-informed neural networks and Physics-informed DeepONet excel in solving partial differential equations; however, they often fail to converge for singularly perturbed problems. To address this, we propose two novel frameworks, Prandtl-Van Dyke neural network(PVD-Net) and its operator learning extension Prandtl-Van Dyke Deep Operator Network (PVD-ONet), which rely solely on governing equations without data. To address varying task-specific requirements, both PVD-Net and PVD-ONet are developed in two distinct versions, tailored respectively for stability-focused and high-accuracy modeling. The leading-order PVD-Net adopts a two-network architecture combined with Prandtl's matching condition, targeting stability-prioritized scenarios. The high-order PVD-Net employs a five-network design with Van Dyke's matching principle to capture fine-scale boundary layer structures, making it ideal for high-accuracy scenarios. PVD-ONet generalizes PVD-Net to the operator learning setting by assembling multiple DeepONet modules, directly mapping initial conditions to solution operators and enabling instant predictions for an entire family of boundary layer problems without retraining. Numerical experiments (second-order equations with constant and variable coefficients, and internal layer problems) show that the proposed methods consistently outperform existing baselines. Moreover, beyond forward prediction, the proposed framework can be extended to inverse problems. It enables the inference of the scaling exponent governing boundary layer thickness from sparse data, providing potential for practical applications.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Prandtl-Van Dyke neural network (PVD-Net) and its DeepONet extension (PVD-ONet) for singularly perturbed boundary-layer problems. Two variants are introduced: a two-network architecture using Prandtl matching for stability and a five-network architecture using Van Dyke matching for higher accuracy. The methods embed classical asymptotic matching conditions as soft penalties in the physics-informed loss and extend to operator learning for mapping initial conditions to solution operators. Numerical experiments on second-order equations (constant and variable coefficients) and internal-layer problems are said to show consistent outperformance over baselines; the framework is also applied to inverse inference of the boundary-layer scaling exponent from sparse data.

Significance. If the central claims are substantiated, the work offers a principled way to incorporate asymptotic structure into PINN/DeepONet training for stiff singularly perturbed problems where standard physics-informed approaches often fail to converge. The operator-learning generalization and the inverse-problem capability are potentially useful for families of problems and parameter identification. Credit is due for grounding the architectures in established matching principles rather than ad-hoc regularization.

major comments (2)
  1. [Section 3.2] Section 3.2 (five-network high-order PVD-Net): the claim that Van Dyke matching as a soft penalty produces accurate composite solutions for variable-coefficient problems requires explicit verification. For variable-coefficient outer expansions the matching relations involve slow-variable-dependent coefficients; it is not shown that the relative weighting of the matching term is sufficient to prevent the optimizer from satisfying PDE residuals while violating higher-order asymptotic consistency inside the layer. A quantitative check (e.g., pointwise deviation from the matched asymptotic expansion at selected orders) is needed to support the reported superiority.
  2. [Numerical experiments] Numerical experiments section: the abstract and introduction assert consistent outperformance and successful inverse inference, yet no error tables, convergence plots with respect to network width/depth, or ablation studies isolating the contribution of the matching penalties are referenced. Without these, the load-bearing claim that the Prandtl/Van Dyke constraints are responsible for the improvement cannot be evaluated.
minor comments (2)
  1. Notation for the composite solution and the inner/outer networks should be introduced once with a clear diagram; repeated re-definition across sections reduces readability.
  2. The description of the two-network versus five-network architectures would benefit from an explicit statement of which matching orders are retained in each case.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive review of our manuscript arXiv:2507.21437. We address each major comment below and describe the revisions we will implement to strengthen the presentation.

read point-by-point responses

  1. Referee: [Section 3.2] Section 3.2 (five-network high-order PVD-Net): the claim that Van Dyke matching as a soft penalty produces accurate composite solutions for variable-coefficient problems requires explicit verification. For variable-coefficient outer expansions the matching relations involve slow-variable-dependent coefficients; it is not shown that the relative weighting of the matching term is sufficient to prevent the optimizer from satisfying PDE residuals while violating higher-order asymptotic consistency inside the layer. A quantitative check (e.g., pointwise deviation from the matched asymptotic expansion at selected orders) is needed to support the reported superiority.

    Authors: We agree that explicit quantitative verification is required to confirm asymptotic consistency for the five-network architecture on variable-coefficient problems. In the revised manuscript we will add a dedicated verification subsection that reports pointwise deviations between the learned composite solution and the matched asymptotic expansion (including slow-variable-dependent coefficients) at the orders retained in the Van Dyke matching. These checks will be performed at representative points inside and outside the layer to demonstrate that the soft penalty enforces the required matching without allowing the optimizer to trade off higher-order consistency against PDE residuals. revision: yes

  2. Referee: [Numerical experiments] Numerical experiments section: the abstract and introduction assert consistent outperformance and successful inverse inference, yet no error tables, convergence plots with respect to network width/depth, or ablation studies isolating the contribution of the matching penalties are referenced. Without these, the load-bearing claim that the Prandtl/Van Dyke constraints are responsible for the improvement cannot be evaluated.

    Authors: We acknowledge that the current numerical section would benefit from more systematic quantitative support. We will expand the experiments section to include (i) comprehensive error tables for all forward and inverse test cases, (ii) convergence plots versus network width and depth, and (iii) ablation studies that systematically vary the weighting of the Prandtl and Van Dyke matching terms while holding all other hyperparameters fixed. These additions will directly isolate the contribution of the asymptotic matching penalties to the observed improvements over baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation rests on classical external matching principles and PDE residuals

full rationale

The paper defines PVD-Net and PVD-ONet by embedding Prandtl's matching condition (two-network version) and Van Dyke's matching principle (five-network version) as soft constraints inside a composite loss that also includes the governing PDE residuals. These matching rules are standard results from the classical asymptotic analysis literature (Prandtl, Van Dyke) and are not derived from or fitted to the present networks' outputs. The operator-learning extension assembles DeepONet modules to map initial conditions to solution operators without introducing any self-referential definition or fitted parameter that is later relabeled as a prediction. Numerical claims are supported by direct comparison against baselines on the same PDEs; no load-bearing step reduces by construction to a self-citation or to a quantity defined in terms of the target result. The framework therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on classical asymptotic matching rules and standard neural-network training assumptions rather than on newly postulated physical entities.

free parameters (1)
  • network topology choice
    Decision to use either a two-network or five-network architecture depending on stability versus accuracy priority.
axioms (1)
  • domain assumption Prandtl matching condition and Van Dyke matching principle accurately glue inner and outer solutions for the boundary-layer problems under study.
    Invoked to construct the composite loss that enforces continuity across the layer.

pith-pipeline@v0.9.0 · 5792 in / 1347 out tokens · 36225 ms · 2026-05-19T03:06:22.142371+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

56 extracted references · 56 canonical work pages · 3 internal anchors

  1. [1]

    F. M. White, J. Majdalani, Viscous fluid flow, Vol. 3, McGraw-Hill New York, 2006

    work page 2006

  2. [2]

    F. M. White, Fluid mechanics 8th edition (2017)

    work page 2017

  3. [3]

    Schoppa, F

    W. Schoppa, F. Hussain, A large-scale control strategy for drag reduction in turbulent boundary layers, Physics of Fluids 10 (5) (1998) 1049–1051. doi:https://doi.org/10.1063/1.869789

    work page doi:10.1063/1.869789 1998

  4. [4]

    J. D. Anderson, Hypersonic and high temperature gas dynamics, AIAA, 1989

    work page 1989

  5. [5]

    Anderson, Fundamentals of Aerodynamics, McGraw Hill, 2011

    J. Anderson, Fundamentals of Aerodynamics, McGraw Hill, 2011

    work page 2011

  6. [6]

    E. L. Reiss, Symmetric bending of thick circular plates, Journal of the Society for Industrial and Applied Mathematics 10 (4) (1962) 596–609. doi:https://doi.org/10.1137/0110045

    work page doi:10.1137/0110045 1962

  7. [7]

    Chien, Large deflection of a circular clamped plate under uniform pressure, Chinese Journal of Physics 7 (2) (1947) 102–113

    W.-Z. Chien, Large deflection of a circular clamped plate under uniform pressure, Chinese Journal of Physics 7 (2) (1947) 102–113

    work page 1947

  8. [8]

    W. E. Alzheimer, R. Davis, Unsymmetrical bending of prestressed annular plates, Journal of the Engineering Mechanics Division 94 (4) (1968) 905–918. doi:https://doi.org/10.1061/JMCEA3.0001000

    work page doi:10.1061/jmcea3.0001000 1968

  9. [9]

    Arzani, A

    A. Arzani, A. M. Gambaruto, G. Chen, S. C. Shadden, Lagrangian wall shear stress structures and near-wall transport in high-schmidt-number aneurysmal flows, Journal of Fluid Mechanics 790 (2016) 158–172. doi:https://doi.org/10.1017/ jfm.2016.6

    work page 2016

  10. [10]

    A. H. Nayfeh, Perturbation methods, John Wiley & Sons, 2024

    work page 2024

  11. [11]

    Prandtl, On fluid motions with very small friction, Verhldg 3 (1904) 484–491

    L. Prandtl, On fluid motions with very small friction, Verhldg 3 (1904) 484–491

    work page 1904

  12. [12]

    Von K´ arm´ an, Uber laminare und turbulente reibung, Z

    T. Von K´ arm´ an, Uber laminare und turbulente reibung, Z. Angew. Math. Mech. 1 (1921) 233–252. doi:https://doi. org/10.1002/zamm.19210010401

    work page doi:10.1002/zamm.19210010401 1921

  13. [13]

    Van Dyke, Higher approximations in boundary-layer theory part 1

    M. Van Dyke, Higher approximations in boundary-layer theory part 1. general analysis, Journal of Fluid Mechanics 14 (2) (1962) 161–177. doi:https://doi.org/10.1017/S0022112062001147

    work page doi:10.1017/s0022112062001147 1962

  14. [14]

    Van Dyke, Higher approximations in boundary-layer theory part 2

    M. Van Dyke, Higher approximations in boundary-layer theory part 2. application to leading edges, Journal of Fluid Mechanics 14 (4) (1962) 481–495. doi:https://doi.org/10.1017/S0022112062001391. 31

    work page doi:10.1017/s0022112062001391 1962

  15. [15]

    Van Dyke, Higher approximations in boundary-layer theory part 3

    M. Van Dyke, Higher approximations in boundary-layer theory part 3. parabola in uniform stream, Journal of Fluid Mechanics 19 (1) (1964) 145–159. doi:https://doi.org/10.1017/S002211206400060X

    work page doi:10.1017/s002211206400060x 1964

  16. [16]

    R. E. O’Malley, A boundary value problem for certain nonlinear second order differential equations with a small parameter, Archive for Rational Mechanics and Analysis 29 (1968) 66–74

    work page 1968

  17. [17]

    R. E. O’Malley, Topics in singular perturbations, Advances in Mathematics 2 (4) (1968) 365–470. doi:https://doi.org/ 10.1016/0001-8708(68)90023-6

    work page doi:10.1016/0001-8708(68)90023-6 1968

  18. [18]

    R. E. O’Malley, Boundary value problems for linear systems of ordinary differential equations involving many small parameters, Journal of Mathematics and Mechanics 18 (9) (1969) 835–855

    work page 1969

  19. [19]

    R. E. O’Malley, Boundary layer methods for nonlinear initial value problems, SIAM Review 13 (4) (1971) 425–434. doi:https://doi.org/10.1137/1013091

    work page doi:10.1137/1013091 1971

  20. [20]

    A. H. Nayfeh, A generalized method for treating singular perturbation problems, Stanford University, 1964

    work page 1964

  21. [21]

    A. H. Nayfeh, A comparison of three perturbation methods for earth-moon-spaceship problem, AIAA Journal 3 (9) (1965) 1682–1687. doi:https://doi.org/10.2514/3.3226

    work page doi:10.2514/3.3226 1965

  22. [22]

    A. H. Nayfeh, A perturbation method for treating nonlinear oscillation problems, Journal of Mathematics and Physics 44 (1-4) (1965) 368–374. doi:https://doi.org/10.1002/sapm1965441368

    work page doi:10.1002/sapm1965441368 1965

  23. [23]

    T. J. Hughes, The finite element method: linear static and dynamic finite element analysis, Courier Corporation, 2012

    work page 2012

  24. [24]

    Eymard, T

    R. Eymard, T. Gallou¨ et, R. Herbin, Finite volume methods, Handbook of Numerical Analysis 7 (2000) 713–1018. doi: https://doi.org/10.1016/S1570-8659(00)07005-8

    work page doi:10.1016/s1570-8659(00)07005-8 2000

  25. [25]

    J. C. Strikwerda, Finite difference schemes and partial differential equations, SIAM, 2004

    work page 2004

  26. [26]

    J. Ling, A. Kurzawski, J. Templeton, Reynolds averaged turbulence modelling using deep neural networks with embedded invariance, Journal of Fluid Mechanics 807 (2016) 155–166. doi:https://doi.org/10.1017/jfm.2016.615

    work page doi:10.1017/jfm.2016.615 2016

  27. [27]

    P. A. Srinivasan, L. Guastoni, H. Azizpour, P. Schlatter, R. Vinuesa, Predictions of turbulent shear flows using deep neural networks, Physical Review Fluids 4 (5) (2019) 054603. doi:https://doi.org/10.1103/PhysRevFluids.4.054603

    work page doi:10.1103/physrevfluids.4.054603 2019

  28. [28]

    Jiang, R

    C. Jiang, R. Vinuesa, R. Chen, J. Mi, S. Laima, H. Li, An interpretable framework of data-driven turbulence modeling using deep neural networks, Physics of Fluids 33 (5) (2021) 055133. doi:https://doi.org/10.1063/5.0048909

    work page doi:10.1063/5.0048909 2021

  29. [29]

    S. L. Brunton, B. R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annual Review of Fluid Mechanics 52 (1) (2020) 477–508. doi:https://doi.org/10.1146/annurev-fluid-010719-060214

    work page doi:10.1146/annurev-fluid-010719-060214 2020

  30. [30]

    Journal of Computational Physics 378, 686–707

    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. doi:https://doi.org/10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  31. [31]

    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. doi:https://doi.org/10.1137/19M1274067

    work page doi:10.1137/19m1274067 2021

  32. [32]

    Arzani, J.-X

    A. Arzani, J.-X. Wang, R. M. D’Souza, Uncovering near-wall blood flow from sparse data with physics-informed neural networks, Physics of Fluids 33 (7) (2021) 071905. doi:https://doi.org/10.1063/5.0055600

    work page doi:10.1063/5.0055600 2021

  33. [33]

    W. Ji, W. Qiu, Z. Shi, S. Pan, S. Deng, Stiff-pinn: Physics-informed neural network for stiff chemical kinetics, The Journal of Physical Chemistry A 125 (36) (2021) 8098–8106. doi:https://pubs.acs.org/doi/10.1021/acs.jpca.1c05102

    work page doi:10.1021/acs.jpca.1c05102 2021

  34. [34]

    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

    work page 2021

  35. [35]

    Z. Mao, A. D. Jagtap, G. E. Karniadakis, Physics-informed neural networks for high-speed flows, Computer Methods in Applied Mechanics and Engineering 360 (2020) 112789. doi:https://doi.org/10.1016/j.cma.2019.112789

    work page doi:10.1016/j.cma.2019.112789 2020

  36. [36]

    C. Rao, H. Sun, Y. Liu, Physics-informed deep learning for incompressible laminar flows, Theoretical and Applied Me- chanics Letters 10 (3) (2020) 207–212. doi:https://doi.org/10.1016/j.taml.2020.01.039

    work page doi:10.1016/j.taml.2020.01.039 2020

  37. [37]

    Shukla, A

    K. Shukla, A. D. Jagtap, J. L. Blackshire, D. Sparkman, G. E. Karniadakis, A physics-informed neural network for 32 quantifying the microstructural properties of polycrystalline nickel using ultrasound data: A promising approach for solving inverse problems, IEEE Signal Processing Magazine 39 (1) (2021) 68–77. doi:10.1109/MSP.2021.3118904

    work page doi:10.1109/msp.2021.3118904 2021

  38. [38]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics-informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440

    work page 2021

  39. [39]

    Discretization of Dirac delta functions in level set methods.Journal of Computational Physics, 207(1):28–51, July 2005.doi:10.1016/j.jcp

    A. Arzani, K. W. Cassel, R. M. D’Souza, Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation, Journal of Computational Physics 473 (2023) 111768. doi:https://doi.org/10.1016/j.jcp. 2022.111768

    work page doi:10.1016/j.jcp 2023

  40. [40]

    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 Computa- tional Physics 28 (5) (2020) 2002–2041. doi:https://doi.org/10.4208/cicp.oa-2020-0164

    work page doi:10.4208/cicp.oa-2020-0164 2020

  41. [41]

    M. A. Nabian, R. J. Gladstone, H. Meidani, Efficient training of physics-informed neural networks via importance sampling, Computer-Aided Civil and Infrastructure Engineering 36 (8) (2021) 962–977. doi:https://doi.org/10.1111/mice.12685

    work page doi:10.1111/mice.12685 2021

  42. [42]

    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. doi:https://doi.org/10.1016/j.cma.2021.113938

    work page doi:10.1016/j.cma.2021.113938 2021

  43. [43]

    Kaplun, P

    S. Kaplun, P. A. Lagerstrom, L. N. Howard, C.-s. Liu, Fluid Mechanics and Singular Perturbations: A Collection of Papers, Academic Press, New York/London, 1967

    work page 1967

  44. [44]

    Zhang, G

    L. Zhang, G. He, Multi-scale-matching neural networks for thin plate bending problem, Theoretical and Applied Mechanics Letters 14 (1) (2024) 100494. doi:https://doi.org/10.1016/j.taml.2024.100494

    work page doi:10.1016/j.taml.2024.100494 2024

  45. [45]

    F. Cao, F. Gao, X. Guo, D. Yuan, Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection-dominated problem, Computers & Mathematics with Applications 150 (2023) 229–242. doi:https://doi.org/10.1016/j.camwa.2023.09.030

    work page doi:10.1016/j.camwa.2023.09.030 2023

  46. [46]

    Huang, R

    J. Huang, R. Qiu, J. Wang, Y. Wang, Multi-scale physics-informed neural networks for solving high reynolds number boundary layer flows based on matched asymptotic expansions, Theoretical and Applied Mechanics Letters 14 (2) (2024) 100496. doi:https://doi.org/10.1016/j.taml.2024.100496

    work page doi:10.1016/j.taml.2024.100496 2024

  47. [47]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Graph kernel network for partial differential equations (2020). arXiv:2003.03485

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  48. [48]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (3) (2021) 218–229

    work page 2021

  49. [49]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations (2020). arXiv:2010.08895

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  50. [50]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to PDEs, Journal of Machine Learning Research 24 (89) (2023) 1–97

    work page 2023

  51. [51]

    T. Du, Z. Huang, Y. Li, Approximation and generalization of DeepONets for learning operators arising from a class of singularly perturbed problems (2023). arXiv:2306.16833

    work page arXiv 2023

  52. [52]

    S. Wang, H. Wang, P. Perdikaris, Learning the solution operator of parametric partial differential equations with physics- informed DeepONets, Science Advances 7 (40) (2021) eabi8605. doi:10.1126/sciadv.abi8605

    work page doi:10.1126/sciadv.abi8605 2021

  53. [53]

    Erd´ elyi, An expansion procedure for singular perturbations, Atti Accad

    A. Erd´ elyi, An expansion procedure for singular perturbations, Atti Accad. Sci. Torino Cl. Sci. Fis. Mat. Nat, 1961

    work page 1961

  54. [54]

    A. G. Baydin, B. A. Pearlmutter, A. A. Radul, J. M. Siskind, Automatic differentiation in machine learning: a survey, Journal of Machine Learning Research 18 (153) (2018) 1–43

    work page 2018

  55. [55]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  56. [56]

    D. C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization, Mathematical Programming 45 (1) (1989) 503–528. 33

    work page 1989