Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Solving all laminar flows around airfoils all-at-once using a parametric neural network solver

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims a single neural network can solve all steady laminar airfoil flows in a given parameter space, in one 4.6-day training run.

desk verdict Real TSONN progress on parametric laminar airfoil flows, but the 'all laminar flows' claim needs a steady-flow qualifier and a high-Re/high-alpha check. read the letter →

arxiv 2501.01165 v1 pith:6UN2AH22 submitted 2025-01-02 physics.flu-dyn

Wenbo Cao , Shixiang Tang , Qianhong Ma , Wanli Ouyang , Weiwei Zhang This is my paper · ORCID
classification physics.flu-dyn
keywords physics-informedneuralnetworkstime-stepping-orientednetworklaminarairfoilflowparametricNavier-Stokesequationssurrogatemodelingmeshtransformationshapeparameterization
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

Steady laminar flow around an airfoil is normally solved one case at a time, with each Reynolds number, angle of attack, and geometry requiring its own CFD run. This paper claims that a single neural network can instead solve the whole parameter space at once: steady incompressible laminar flows for Reynolds numbers 100–5000, angles of attack −5° to 15°, and airfoil shapes obtained by 30% perturbations of the UIUC database. The method combines a time-stepping-oriented neural network (TSONN), which cures the ill-conditioning of physics-informed networks, with mesh transformation, so that shape enters only through metric coefficients. The authors report a 4.6-day training run, about 40 times the cost of a single flow, producing a surrogate with mean relative errors of 4.6% in lift and 1.1% in drag. If this holds, a designer gets near-instant predictions across a wide design space without building a labeled CFD dataset.

What carries the argument

The machinery has two main pieces. TSONN replaces the direct minimization of the steady PDE residual with a sequence of implicit pseudo-time steps: at outer iteration $n$ the network is trained to satisfy a combination of the residual and the pseudo-time increment $(q_{n+1}-q_n)/\Delta\tau$, converting one poorly conditioned optimization into many well-conditioned ones solved with LBFGS. Mesh transformation maps the physical domain around the airfoil to a uniform rectangular computational coordinate system, with the shape encoded in the metric coefficients; mesh deformation via radial basis function interpolation produces coordinates and metrics for each perturbed shape per collocation point. The third ingredient, volume-weighted PDE residuals, reweights collocation points by local cell volume so the dense wall region drives the loss appropriately. Each batch contains 30,000 interior and 1,000 boundary points with randomly sampled Reynolds number, angle of attack, and shape, so one 10,000-iteration run visits $3.1\times10^8$ distinct collocation points while using no labeled flow data.

What would settle it

Run a time-resolved simulation or experiment at a point inside the parameter space, say NACA0012 at Re=5000 and angle of attack 15 degrees, and check whether periodic vortex shedding appears; if it does, the steady surrogate cannot be solving the real laminar flow there, which settles that the 'all laminar flows' claim is false in that region.

Watch

Extended reading notes

Core claim

The central discovery, as the authors state it, is that the ill-conditioning that has limited physics-informed neural networks for airfoil flows can be bypassed by pseudo-time stepping, and that the resulting solver extends to high-dimensional parametric problems essentially for free. The parametric model takes the computational-space coordinates, Reynolds number, angle of attack, and a shape parameter vector as inputs and outputs velocity and pressure fields. During training, each collocation point is assigned a random condition and shape sampled from the parameter space, so hundreds of millions of distinct cases are encountered over 10,000 outer iterations. The trained surrogate then gives the flow field, wall pressure, skin friction, and integrated lift and drag for any case in the space. The paper's quantitative claim is that this surrogate matches reference finite-volume solutions with mean relative errors of 4.6% for lift and 1.1% for drag, essentially the same accuracy as solving a single flow, while costing about 40 times one flow to train.

Load-bearing premise

The load-bearing premise is that every point in the stated parameter space actually has a meaningful steady laminar solution to the incompressible Navier-Stokes equations; if parts of that space are unsteady, the steady solver is solving equations that do not represent the real flow there.

Editorial extensions

If this is right

  • A single trained model can be queried in near real time for flow fields, lift, and drag anywhere inside the parameter space, eliminating the need to run a new CFD case for each design point.
  • The total compute of 4.6 days at roughly 40 times a single-flow cost places the whole parameter-space solve within reach of one workstation GPU, making the approach practical for preliminary airfoil design.
  • Because the 204-dimensional y-coordinate shape representation converges like the 16-dimensional CST representation, the solver appears insensitive to how shapes are parameterized, which points toward free-form shape optimization without a fixed parameterization.
  • The surrogate's accuracy (4.6% lift, 1.1% drag) is close to that of the single-flow solver, suggesting that parametric training does not cause a large accuracy penalty.
  • Expanding the sampled parameter space during training (Model 3) extends valid predictions beyond the original range, so the usable domain of the surrogate can be deliberately set by choosing the training space.

Reading between the lines

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

  • The 'all laminar flows' claim is implicitly limited to steady flows; at the high-Reynolds end of the range unsteady vortex shedding is physical, so the surrogate likely misrepresents those corners of the space.
  • The training trick of assigning each collocation point a random condition and shape is a transferable recipe; other parametric PDE solvers could adopt it to cover large spaces without labeled data.
  • The similar convergence of the 16- and 204-dimensional shape representations suggests the network discovers low-dimensional geometric features internally, pointing toward free-form shape optimization without a fixed parameterization.
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

4 major / 5 minor

Summary. The paper proposes a parametric solver for two-dimensional steady incompressible laminar airfoil flows based on the time-stepping-oriented neural network (TSONN) and mesh transformation. The solver maps a computational-space mesh onto the airfoil geometry and takes as inputs the computational coordinates, Reynolds number, angle of attack, and a shape parameter vector, with the shape space built from perturbed CST parameters or y-coordinates of airfoils in the UIUC database. After validating the solver on six single-flow cases and comparing it against a standard PINN, the authors train three parametric surrogate models over parameter spaces that include Re in [10, 5000] (or wider) and angles of attack in [-5, 15] degrees (or wider). They report mean relative errors of 3.6% for lift and 1.4% for drag in the single-flow cases, and 4.6% for lift and 1.1% for drag for the parametric surrogate, with a training cost of 4.6 days, about 40 times the cost of a single-flow solve. The code and models are publicly released.

Significance. If the parameter-space claim can be established, this is a practically useful contribution: a single trained network could replace many CFD runs for steady laminar airfoil flows, and the TSONN formulation provides a concrete remedy for the ill-conditioning that limits standard PINNs on this class of problems. The mesh-transformation and volume-weighted residual techniques are sensible, and the comparison with a standard PINN under identical settings is informative. The public release of the surrogate models and code is a notable strength. The principal caveat is that the central claim of solving 'all laminar flows' in the stated parameter range presumes the existence and physical relevance of steady solutions everywhere in that range; the current evidence, based on only six validation cases that avoid the most problematic high-Reynolds/high-angle region, does not yet establish that presumption.

major comments (4)
  1. [Abstract; §3.3; Eq. (1)] The central claim that the solver covers 'all laminar flows' in Re∈[10,5000] and α∈[-5°,15°] presumes that a meaningful steady laminar solution exists throughout this parameter space. Equation (1) is the steady incompressible Navier–Stokes system, and both TSONN and the reference FVM target steady solutions. At Reynolds numbers near 5000 and angles of attack near 10–15°, laminar boundary-layer separation and vortex shedding are common, so the steady solution may not exist, may be unstable, or may not represent the physically realized flow. Table 1 never combines high Re with high α (the most stressing pairs are Re=5000, α=-5° and Re=100, α=15°), so the six validation cases cannot establish the claim in this region. Please either explicitly restrict the claim to steady laminar solutions, or provide evidence (e.g., unsteady simulations or linear-stability checks) that unsteadiness is absent throughout the stated parameter space, and add validation cases in the high-Re/high-α corner.
  2. [§3.3; Tables 2 and 3] The headline accuracy statistics rest on only six in-distribution validation cases, and the lift mean is computed over five cases because Case 5 is excluded (as disclosed) due to its small reference value. This sample is too small and too selective to support the claims of 'all laminar flows' or 'strong generalization capability.' Please add a larger validation set drawn randomly from the parameter space, including the high-Re/high-α corner, report per-case errors with confidence intervals over multiple training seeds, and make the exclusion of Case 5 from the headline lift error explicit in the abstract and conclusions.
  3. [§3.3; Figure 13] The out-of-distribution generalization comparison for Model 3 is encouraging, but it is demonstrated on only one airfoil (NACA2412) at a single Reynolds number (Re=500). The text states that this demonstrates the capability to 'arbitrarily extend' the model generalization space, which is not supported by a single example. Please either temper this claim or test Model 3 on additional airfoils and Reynolds numbers, both inside and outside its training range.
  4. [§2.4] The statement that training involves 'hundreds of millions of flow conditions and airfoil shapes' should be qualified. The procedure samples 31,000 collocation points per outer iteration for 10,000 outer iterations, giving 3.1×10^8 point-condition samples; these are not hundreds of millions of distinct flow solutions or even distinct converged flow fields. The statement is not incorrect, but as written it may mislead readers into thinking the model was trained on a dataset of solved flows. Please rephrase to clarify that these are residual-evaluation samples.
minor comments (5)
  1. [§3.3, tables/model list] The text says 'as shown in Table 1, Model 1 and Model 2 represent shapes...' but Table 1 lists the six validation cases, not the model configurations. This cross-reference is broken; it should refer to the correct table listing Models 1–3.
  2. [§3.2 and §3.3, Case 5] Since Case 5 is excluded from the mean relative lift error, the relative error for Case 5 should be reported explicitly (for the single-flow case it is about 36%, and for Model 1 about 14%), so that readers can see the magnitude that the mean is hiding.
  3. [Equations in §2.1–§2.3] Several equations, especially Eq. (1), Eq. (2), and Eq. (5), appear garbled in the manuscript text (e.g., missing vector symbols and misplaced operators). Please ensure the final typeset version has correct mathematical notation.
  4. [§3.3, Figure 8] The caption of Figure 8 says 'Error convergence curves for solving parametric problem and a single flow,' but it does not state which single-flow case is plotted or that the error metric is the average relative L1 error of the wall pressure coefficient. Please add this information to the caption.
  5. [Abstract and §3.3] Minor language issues: 'in just 4.6 day' should be '4.6 days'; 'covering nearly all laminar flow scenarios in engineering' and 'arbitrarily extend the model generalization space' are stronger than the evidence supports and should be softened or qualified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the solver is trained against PDE residuals and independently benchmarked against FVM references.

full rationale

The paper's claimed derivation chain is self-contained in the relevant sense. The target quantities (velocity, pressure, lift, drag) are obtained by minimizing PDE and boundary residuals in Equations (3)-(5), not by fitting to the reference lift/drag coefficients reported in Tables 2 and 3. The reference FVM solutions are generated by an independent in-house solver described in Section 3, and the reported errors are computed against those references, so the accuracy claims are not forced by construction. Self-citations to the authors' prior TSONN work [21,25], mesh-transformation papers [20,22,26], and volume weighting [27] are methodological building blocks; their effectiveness is re-demonstrated in this paper through direct comparison with PINNs and FVM (Figures 4-6, Tables 2-3), so they are not load-bearing circular appeals. The 'all laminar flows' language is a scope assertion about the parameter space, which is defined independently of the output quantities; concerns about whether steady solutions exist at high Reynolds number and high angle of attack are external-validity or correctness issues, not circularity. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported solely from the authors' prior work. The evaluation on held-out cases from the UIUC database provides independent empirical support for the generalization claim.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-selected hyperparameters and several domain assumptions, most notably steadiness of the flow, that are not stress-tested. No new physical entity is introduced.

free parameters (6)
  • PDE loss weight lambda_PDE = 100
    Relative weight of PDE residual to boundary residual; chosen by hand and validated in prior TSONN papers, not by sensitivity study here.
  • Boundary loss weight lambda_BC = 1
    Fixed boundary condition weight; held constant across all cases.
  • Pseudo-time step Delta_tau = 0.3
    Controls TSONN stability and accuracy; chosen from previous studies and not tuned per case.
  • Inner optimizer iterations K = 500
    LBFGS iterations per TSONN outer step; hand-selected.
  • Network architecture = 6 or 10 hidden layers, 128 neurons per layer, tanh
    Wider and deeper network for parametric problems; chosen by hand, no architecture ablation.
  • CST perturbation amplitude = ±30% (Models 1/2: ±35%; Model 3: ±50%)
    Defines the claimed shape coverage; arbitrary choice not grounded in physical coverage analysis.
assumptions (5)
  • domain assumption Steady incompressible Navier-Stokes equations correctly model all flows in the parameter space.
    Invoked in Eq. (1); ignores unsteadiness at higher Re and alpha.
  • domain assumption Every point in the parameter space has a unique steady solution reachable by pseudo-time stepping.
    TSONN and FVM both assume convergence to steady state; unstable or bifurcating flows are excluded.
  • domain assumption RBF mesh deformation produces valid, smooth meshes for every perturbed airfoil in the parameter space.
    Section 2.4; invalid meshes would corrupt metric coefficients and the PDE residual.
  • domain assumption CST with 12 parameters represents UIUC airfoils and their perturbations well.
    Section 2.4; approximation error is not quantified.
  • domain assumption The in-house finite volume solver is an accurate reference.
    Section 3; no grid-convergence or comparison to external benchmark is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solving all laminar flows around airfoils all-at-once using a parametric neural network solver." pith.science (2026). https://pith.science/paper/6UN2AH22

@misc{pith2026250101165,
  author       = {Pith},
  title        = {Pith review of: Solving all laminar flows around airfoils all-at-once using a parametric neural network solver},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UN2AH22}},
  note         = {Machine review of arXiv:2501.01165}
}
read the original abstract

Recent years have witnessed increasing research interests of physics-informed neural networks (PINNs) in solving forward, inverse, and parametric problems governed by partial differential equations (PDEs). Despite their promise, PINNs still face significant challenges in many scenarios due to ill-conditioning. Time-stepping-oriented neural network (TSONN) addresses this by reformulating the ill-conditioned optimization problem into a series of well-conditioned sub-problems, greatly improving its ability to handle complex scenarios. This paper presents a new solver for laminar flow around airfoils based on TSONN and mesh transformation, validated across various test cases. Specifically, the solver achieves mean relative errors of approximately 3.6% for lift coefficients and 1.4% for drag coefficients. Furthermore, this paper extends the solver to parametric problems involving flow conditions and airfoil shapes, covering nearly all laminar flow scenarios in engineering. The shape parameter space is defined as the union of 30% perturbations applied to each airfoil in the UIUC airfoil database, with Reynolds numbers ranging from 100 to 5000 and angles of attack spanning from -5{\deg} to 15{\deg}. The parametric solver solves all laminar flows within the parameter space in just 4.6 day, at approximately 40 times the computational cost of solving a single flow. The model training involves hundreds of millions of flow conditions and airfoil shapes, ultimately yielding a surrogate model with strong generalization capability that does not require labeled data. Specifically, the surrogate model achieves average errors of 4.6% for lift coefficients and 1.1% for drag coefficients, demonstrating its potential for high generalizability, cost-effectiveness, and efficiency in addressing high-dimensional parametric problems and surrogate modeling.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FENN: Feature-enhanced neural network for solving partial differential equations involving fluid mechanics

    physics.flu-dyn 2025-01 conditional novelty 6.0 of 10

    Adding distance-angle or potential-flow features to PINN inputs, with offline feature networks to keep derivatives correct, cuts training cost about fourfold while improving accuracy.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in neural information processing systems, 25 (2012)

    work page 2012

  2. [2]

    T.N. Kipf, M. Welling, Semi -supervised classification with graph convolutional networks, arXiv preprint arXiv:1609.02907, (2016)

    arXiv 2016

  3. [3]

    Sutskever, O

    I. Sutskever, O. Vinyals, Q.V . Le, Sequence to sequence learning with neural networks, Advances in neural information processing systems, 27 (2014)

    work page 2014

  4. [4]

    Jaderberg, K

    M. Jaderberg, K. Simonyan, A. Zisserman, Spatial transformer networks, Advances in neural information processing systems, 28 (2015)

    work page 2015

  5. [5]

    J. Hu, W. Zhang, Mesh -Conv: Convolution operator with mesh resolution independence for flow field modeling, Journal of Computational Physics, 452 (2022)

    work page 2022

  6. [6]

    Shukla, V

    K. Shukla, V . Oommen, A. Peyvan, M. Penwarden, N. Plewacki, L. Bravo, A. Ghoshal, R.M. Kirby, G.E. Karniadakis, Deep neural operators as accurate surrogates for shape optimization, Engineering Applications of Artificial Intelligence, 129 (2024)

    work page 2024

  7. [7]

    Zhang, X

    W. Zhang, X. Peng, J. Kou, X. Wang, Heterogeneous data-driven aerodynamic modeling based on physical feature embedding, in, Elsevier, 2024, pp. 1-6

    work page 2024

  8. [8]

    Jordan, T.M

    M.I. Jordan, T.M. Mitchell, Machine learning: Trends, perspectives, and prospects, Science, 349 (2015) 255-260

    work page 2015

Show all 29 references
  1. [9]

    J. Li, X. Du, J.R. Martins, Machine learning in aerodynamic shape optimization, 21 Progress in Aerospace Sciences, 134 (2022) 100849

    2022

  2. [10]

    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

  3. [11]

    Weinan, B

    E. Weinan, B. Yu, The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems, Communications in Mathematics and Statistics, 6 (2018) 1-12

    2018

  4. [12]

    Sirignano, K

    J. Sirignano, K. Spiliopoulos, DGM: A deep learning algorithm for solving partial differential equations, Journal of Computational Physics, 375 (2018) 1339-1364

    2018

  5. [13]

    Karniadakis, I.G

    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

  6. [14]

    Samaniego, C

    E. Samaniego, C. Anitescu, S. Goswami, V .M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang, T. Rabczuk, An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and appl ications, Computer Metho...

    2020

  7. [15]

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

    2021

  8. [16]

    Mao, A.D

    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)

    2020

  9. [17]

    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

  10. [18]

    Wang, H.W

    S.F. Wang, H.W. Wang, P. Perdikaris, Learning the solution operator of parametric partial differential equations with physics -informed DeepONets, Science Advances, 7 (2021) eabi8605

    2021

  11. [19]

    Y . Sun, U. Sengupta, M. Juniper, Physics -informed deep learning for simultaneous surrogate modeling and PDE -constrained optimization of an airfoil geometry, Computer Methods in Applied Mechanics and Engineering, 411 (2023) 116042

    2023

  12. [20]

    W. Cao, J. Song, W. Zhang, Solving high-dimensional parametric engineering problems for inviscid flow around airfoils based on physics-informed neural networks, 22 Journal of Computational Physics, 516 (2024) 113285

    2024

  13. [21]

    W. Cao, W. Zhang, An analysis and solution of ill -conditioning in physics - informed neural networks, Journal of Computational Physics, 520 (2025) 113494

    2025

  14. [22]

    W. Cao, X. Shan, S. Tang, W. Ouyang, W. Zhang, Solving parametric high - Reynolds-number wall -bounded turbulence around airfoils governed by Reynolds - averaged Navier -Stokes equations using time -stepping-oriented neural network, Physics of Fluids, 31 (2025)

    2025

  15. [23]

    Anderson, EBOOK: Fundamentals of Aerodynamics (SI units), McGraw hill, 2011

    J. Anderson, EBOOK: Fundamentals of Aerodynamics (SI units), McGraw hill, 2011

    2011

  16. [24]

    Baydin, B.A

    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 (2018) 1-43

    2018

  17. [25]

    W. Cao, W. Zhang, TSONN: Time -stepping-oriented neural network for solving partial differential equations, arXiv preprint arXiv:2310.16491, (2023)

    2023 arXiv

  18. [26]

    W. Cao, J. Song, W. Zhang, A solver for subsonic flow around airfoils based on physics-informed neural networks and mesh transformation, Physics of Fluids, 36 (2024)

    2024

  19. [27]

    J. Song, W. Cao, F. Liao, W. Zhang, VW-PINNs: A volume weighting method for PDE residuals in physics -informed neural networks, Acta Mechanica Sinica, 41 (2025) 324140

    2025

  20. [28]

    Kulfan, Universal Parametric Geometry Representation Method, Journal of Aircraft, 45 (2008) 142-158

    B.M. Kulfan, Universal Parametric Geometry Representation Method, Journal of Aircraft, 45 (2008) 142-158

    2008

  21. [29]

    De Boer, M.S

    A. De Boer, M.S. Van der Schoot, H. Bijl, Mesh deformation based on radial basis function interpolation, Computers & structures, 85 (2007) 784-795

    2007

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.