Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Geometry-aware PINNs for Turbulent Flow Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A signed-distance geometry encoding lets a physics-informed neural network predict turbulent flow around NACA airfoils at Reynolds numbers absent from training.

desk verdict A plausible geometry-aware RANS-PINN extension with a useful ablation, but an unlisted training set and a sign error in the turbulence equation keep the generalization claim from being convincing. read the letter →

arxiv 2412.01954 v1 pith:VBB3ZRUL submitted 2024-12-02 cs.LG cs.NAmath.NAphysics.flu-dyn

Shinjan Ghosh , Julian Busch , Georgia Olympia Brikis , Biswadip Dey This is my paper · ORCID
classification cs.LGcs.NAmath.NAphysics.flu-dyn
keywords physics-informedneuralnetworksReynolds-averagedNavier-Stokesk-epsilonturbulencemodelsigneddistancefunctionNACA4-digitairfoilsturbulentflowsurrogategeometrygeneralizationparametricPINN
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 claims that a physics-informed neural network can act as a fast surrogate for turbulent flow simulations over airfoils whose shape and Reynolds number it has not seen during training. The key move is to feed the network a signed distance field that describes the airfoil geometry point-by-point, along with the standard NACA 4-digit design numbers and the inlet velocity. The network is trained with a handful of CFD solutions from eight airfoils and supervised by the Reynolds-averaged Navier-Stokes (RANS) equations with a two-equation k-epsilon turbulence model. On held-out NACA airfoils at unseen Reynolds numbers, the predicted velocity and pressure fields have mean errors of roughly 3-6%, with the largest deviations in the thin near-wall boundary layer. If this holds, design exploration could replace repeated costly CFD runs with near-real-time surrogate predictions.

What carries the argument

The load-bearing object is the signed distance function (SDF): a scalar field giving the distance from each point in the domain to the nearest point on the airfoil surface, with zero on the surface. It lets a single network represent different geometries in a mesh-independent way and makes the no-slip boundary condition natural ($U=0$ where the SDF is zero). The global NACA 4-digit parameters (maximum camber, camber position, thickness) and the inlet velocity enter as additional inputs to parameterize shape and Reynolds number. The objective combines sparse CFD data with the RANS conservation equations and the two-equation k-epsilon turbulence model, and a warm-start phase trains on data before turning on the PDE losses.

What would settle it

Run the same trained architecture on a NACA 4-digit airfoil and Reynolds number at the extremes of the parameter space (for example, thickness near 30%, camber position at an extreme, or Reynolds number below 200k or above 700k) and compare the output to a fresh CFD simulation; if the mean velocity error jumps well beyond the 3-6% range seen in the paper, the reported generalization is interpolation within a narrow training envelope rather than a general geometric capability.

Watch

Extended reading notes

Core claim

The paper's central claim is that local geometry information, encoded as a signed distance function, is what enables a parametric RANS-PINN (a physics-informed neural network trained with Reynolds-averaged Navier-Stokes equations) to generalize to airfoil shapes and Reynolds numbers it has not seen. The network takes coordinates, SDF values, the three NACA 4-digit design parameters, and inlet velocity as inputs, and outputs velocity, pressure, turbulent kinetic energy, and dissipation. RANS momentum and continuity equations plus the two-equation k-epsilon closure act as physics losses alongside a small amount of CFD data from eight airfoils, with a warm-start schedule that fits data first and then turns on physics. On three held-out NACA airfoils at unseen Reynolds numbers, the SDF-only model reports mean velocity errors of 3.4-6.0% and mean pressure errors of 1.8-5.1%, with the largest discrepancies concentrated in the near-wall boundary layer. The model using only global design parameters performs worse near the wall and forms phantom-airfoil artifacts, while combining the two geometry channels nearly matches the SDF-only model and improves far-field accuracy.

Load-bearing premise

The load-bearing premise is that the eight training airfoils and six inlet velocities between 2 and 7 m/s adequately span the NACA 4-digit design space and the Reynolds number range of 200k to 700k, so that the held-out validation cases are interpolations inside the training distribution; the paper provides no coverage analysis or distance-to-training-data measure for the validation cases.

Editorial extensions

If this is right

  • Within the trained Reynolds range, design studies could query velocity and pressure fields for new NACA 4-digit airfoils in near real time instead of launching a new CFD solve at every design iteration.
  • The SDF-only model outperforming the global-parameter model in the near-wall region indicates that local geometry encoding carries most of the generalization for unseen shapes; the global parameters are useful but secondary.
  • Because the physics losses supply much of the constraint, the method achieves competitive accuracy with only eight airfoils and six inlet velocities, suggesting data-efficient surrogate construction for other shape families.
  • The combined local-plus-global model nearly matches the best local model and improves far-field behavior, so architectures that fuse local and global geometry information are a promising direction for further accuracy gains.

Reading between the lines

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

  • The paper does not test extrapolation beyond the training envelope; a plausible inference is that the surrogate behaves as an interpolant in the space of camber, camber position, thickness, and Reynolds number, and errors should be expected to grow outside those ranges.
  • If SDF alone generalizes this well within the NACA family, the same local encoding might transfer to other airfoil families or topologies, but the global NACA parameters would have to be replaced by a more general geometric descriptor.
  • The systematic near-wall error suggests a natural next experiment: add a boundary-layer-aware correction or adaptive sampling near the wall, and test whether the mean errors drop below the current 3-6% band.
  • A direct practical extension is to embed the surrogate in an optimizer that screens many airfoil designs cheaply, reserving CFD for the few designs that survive the screening.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 physics-informed neural network (PINN) surrogate for steady turbulent flow over NACA 4-digit airfoils. The model takes spatial coordinates, inlet velocity/Reynolds number, and geometry information (signed distance field and/or NACA design parameters) as inputs, and predicts velocity, pressure, and turbulent quantities (k, epsilon). The loss combines CFD data from 8 airfoils with RANS plus k-epsilon PDE residuals. The authors validate on six held-out cases (Table 1) and report normalized velocity errors mostly below 6%.

Significance. If the results hold, the proposed SDF-based parametric RANS-PINN would be a valuable fast surrogate for airfoil design exploration, extending prior fixed-geometry PINNs to novel shapes. The paper conducts held-out validation on CFD cases, which is appropriate for assessing generalization. The explicit use of signed distance fields as a local geometry embedding is a useful idea. However, the current manuscript contains a technical error in one governing equation and lacks essential details about the training distribution and reproducibility, which substantially weakens the strength of the claims as they stand.

major comments (4)
  1. [Section 2, Eq. (4)] The epsilon equation is written with a plus sign before the C2 epsilon term: (C1 Pk + C2 epsilon) epsilon/k. The standard k-epsilon model has the dissipation term as -C2 rho epsilon^2/k, so the source should be (C1 Pk - C2 epsilon) epsilon/k. As written, the PDE loss enforces an unphysical turbulence model, undermining the claim that the RANS-PINN embeds the correct k-epsilon equations. The authors should correct the sign and, if the reported results were obtained with the erroneous equation, retrain and revalidate the model.
  2. [Section 2, 'A data set comprising of 8 NACA airfoils...'] The eight training airfoils are never listed. Without this list, it is impossible to determine whether the validation airfoils in Table 1 are genuinely unseen, whether they lie inside the parameter space spanned by the training set (camber, camber position, thickness), or whether NACA-1412 (which appears twice in Table 1) is also in the training set. The absence of a coverage analysis (e.g., nearest training geometry distance or convex hull inclusion) means the reported 3-6% errors cannot be attributed to geometric generalization rather than interpolation. The authors should provide the list of training airfoils and a quantitative measure of distance from each validation case to the training set.
  3. [Section 3, Table 1] Quantitative validation consists of six cases, each reported as a single run with no standard deviation or confidence intervals, and all six Reynolds numbers (210k-703k) lie within the training range of 200k-700k. The paper should provide results from multiple training seeds and, to support the abstract's claim of predicting flow at 'any Reynolds number' within the range, demonstrate interpolation behavior more systematically, including at least one case clearly inside the training distribution with a known distance to training data.
  4. [Section 2 and Figure 1] The manuscript does not specify the neural network architecture (depth, width, activation), loss weighting between data and PDE terms, number and distribution of collocation points, SDF discretization, or the warm-start schedule. These details are essential for reproducibility and for assessing whether the reported accuracy is robust to standard hyperparameter choices. The authors should add a full experimental setup section or table.
minor comments (5)
  1. [Throughout] The divergence operator is written as ∇(U) rather than ∇·U; please use consistent vector notation for all divergence terms.
  2. [Section 3.1] The text describes 'the 4 digits from each NACA airfoil' as design inputs, but a NACA 4-digit airfoil is defined by three independent parameters (camber, camber position, thickness); the fourth digit is part of the thickness designation. Please clarify.
  3. [Table 1] Table 1 lists NACA-1412 twice; please either label them as distinct Reynolds numbers (290k and 389k) or rename the cases for clarity.
  4. [Section 4] The conclusion contains an incomplete sentence: 'the current PINN surrogate is able to predict flows over unknown for various NACA airfoils...' Please revise.
  5. [Abstract] The paper claims 'near real-time' prediction in the abstract but does not quantify inference time or the speedup relative to CFD; please add a brief runtime comparison.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the geometry-aware surrogate is validated on held-out CFD cases it never trained on, and the two self-citations (warm-start schedule, prior RANS-PINN baseline) are procedural rather than load-bearing.

full rationale

The paper's central claim is an empirical surrogate result, not a derivation: a PINN embeds geometry via SDF values and NACA parameters plus inlet velocity, is trained against STAR-CCM+ RANS-CFD data from 8 airfoils at 6 inlet velocities, and is evaluated on 6 validation cases (Table 1) that the paper explicitly describes as 'validated with unknown NACA airfoils at unseen Reynolds numbers'. The reported 3-6% velocity errors are computed against external CFD ground truth using formulas (6) and (7), rather than against training data or against quantities defined in terms of the predictions, so they do not reduce to the model's inputs by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, no known result is re-labeled as unification, and the PDE losses are standard RANS/k-epsilon constraints rather than an ansatz smuggled in via citation. The two self-citations to Ghosh et al. [2023] — the warm-start pre-training schedule in Section 2 ('similar to Ghosh et al. [2023]') and the parametric RANS-PINN architecture reused as the G ablation in Section 3.1 ('from previous literature [Ghosh et al., 2023]') — are procedural or baseline references; neither is invoked as evidence for the generalization claim, whose support is the empirical validation against CFD. They are therefore minor self-citations, not load-bearing circularity, consistent with a score of 1. Remaining concerns about the paper — the 8 training airfoils are never listed, so the coverage of the validation cases and a possible overlap with NACA-1412 (listed twice in Table 1) cannot be audited, and Eq. (4) shows a sign anomaly in the epsilon source term relative to the standard k-epsilon model — are reproducibility and correctness risks, not circularity reductions, and are outside the circularity score under the analysis rules.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central result does not introduce new physical entities; it rests on standard turbulence modeling assumptions and unreported training hyperparameters. The free parameters are the hand-chosen loss weights and architecture choices that are not disclosed.

free parameters (3)
  • PDE/data loss weights and warm-start schedule = not reported
    The balance between RANS/k-epsilon residuals and CFD data loss determines the training outcome; the paper neither reports the weights nor the warm-start transition point (Section 2, warm start paragraph), so the results depend on unreported hand-chosen hyperparameters.
  • Neural network architecture and size = not reported
    Figure 1 shows a modular network but gives no layer widths, depth, or activation functions; model capacity is a free choice that affects accuracy.
  • Collocation point count and surface sampling density = not reported
    The physics loss is evaluated at collocation points whose number and distribution are not specified, so the effective PDE constraint is undefined.
assumptions (5)
  • domain assumption RANS equations with an eddy-viscosity closure accurately describe the mean turbulent flow for the airfoil cases
    Used as the governing equations (Eqs. 1-2); validity for airfoil flows with separation is assumed, not verified.
  • domain assumption The k-epsilon model with constants C1=1.44, C2=1.92, sigma_k=1, sigma_eps=1.3 is suitable for this flow regime
    Standard empirical constants are imported without validation for these specific geometries and Reynolds numbers.
  • domain assumption STAR-CCM+ CFD solutions with prism layers and resolved wall y+ are accurate ground truth
    Stated in Section 2; no mesh-convergence or solver-validation study is provided, so the reference data quality is assumed.
  • ad hoc to paper The 8 training airfoils and 6 training velocities cover the NACA 4-digit parameter space and Re range
    The generalization to 'any' unseen airfoil and 'any' Re requires the sparse training set to span the input space; no coverage analysis is given.
  • domain assumption The SDF field plus coordinates uniquely and sufficiently encodes the local geometry for flow prediction
    The L model relies on this representation; the paper does not analyze its completeness or sensitivity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometry-aware PINNs for Turbulent Flow Prediction." pith.science (2026). https://pith.science/paper/VBB3ZRUL

@misc{pith2026241201954,
  author       = {Pith},
  title        = {Pith review of: Geometry-aware PINNs for Turbulent Flow Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VBB3ZRUL}},
  note         = {Machine review of arXiv:2412.01954}
}
abstract

Design exploration or optimization using computational fluid dynamics (CFD) is commonly used in the industry. Geometric variation is a key component of such design problems, especially in turbulent flow scenarios, which involves running costly simulations at every design iteration. While parametric RANS-PINN type approaches have been proven to make effective turbulent surrogates, as a means of predicting unknown Reynolds number flows for a given geometry at near real-time, geometry aware physics informed surrogates with the ability to predict varying geometries are a relatively less studied topic. A novel geometry aware parametric PINN surrogate model has been created, which can predict flow fields for NACA 4 digit airfoils in turbulent conditions, for unseen shapes as well as inlet flow conditions. A local+global approach for embedding has been proposed, where known global design parameters for an airfoil as well as local SDF values can be used as inputs to the model along with velocity inlet/Reynolds number ($\mathcal{R}_e$) to predict the flow fields. A RANS formulation of the Navier-Stokes equations with a 2-equation k-epsilon turbulence model has been used for the PDE losses, in addition to limited CFD data from 8 different NACA airfoils for training. The models have then been validated with unknown NACA airfoils at unseen Reynolds numbers.

Figures

Figures reproduced from arXiv: 2412.01954 by the authors.

Figure 1
Figure 1. Architecture Diagram. A CFD model with wake refinement and prism layers, with all wall y + has been created in STAR￾CCM+ for each NACA airfoil. A velocity inlet boundary condition along with pressure outlets, no-slip walls and free stream side walls have been considered. A data set comprising of 8 NACA airfoils with, six inlet velocities(|U|in) randomly generated have been selected from the range of 2 m/s to 7 m/s, … view at source ↗
Figure 2
Figure 2. Velocity predictions for various NACA airfoils at different Re. Pred Pred Pred True True True Error Error Error NACA-1412, Re=290k NACA-6412, Re=441k NACA-4421, Re=703k 𝑝 (𝑃𝑎) 𝑝 (𝑃𝑎) 𝑝 (𝑃𝑎) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Pressure predictions for various NACA airfoils at different Re. better at the outer zone. This can be attributed to lack of local information which the SDF provides in embedding the knowledge of geometric boundaries [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Prediction comparison of speed for L model and G model. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

  1. [1]

    Chen and N

    L.-W. Chen and N. Thuerey. Towards high-accuracy deep learning inference of compressible flows over aerofoils. Computers and Fluids, 250: 0 105707, 2023. ISSN 0045-7930. doi:https://doi.org/10.1016/j.compfluid.2022.105707. URL https://www.sciencedirect.com/science/article/pii/S0045793022003000

    arXiv 2023

  2. [2]

    Dwivedi, N

    V. Dwivedi, N. Parashar, and B. Srinivasan. Distributed physics informed neural network for data-efficient solution to partial differential equations. arXiv preprint arXiv:1907.08967, 2019

    arXiv 1907

  3. [3]

    Eivazi, M

    H. Eivazi, M. Tahani, P. Schlatter, and R. Vinuesa. Physics-informed neural networks for solving reynolds-averaged navier--stokes equations. Physics of Fluids, 34 0 (7): 0 075117, 2022

    work page 2022

  4. [4]

    Ghosh, A

    S. Ghosh, A. Chakraborty, G. O. Brikis, and B. Dey. RANS-PINN based Simulation Surrogates for Predicting Turbulent Flows . ICML Workshop on Synergy of Scientific and Machine Learning Modeling Workshop (SynS & ML), 2023

    work page 2023

  5. [5]

    X. Guo, W. Li, and F. Iorio. Convolutional neural networks for steady flow approximation. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, page 481–490, New York, NY, USA, 2016. Association for Computing Machinery

    work page 2016

  6. [6]

    Hanrahan, M

    S. Hanrahan, M. Kozul, and R. Sandberg. Studying turbulent flows with physics-informed neural networks and sparse data. International Journal of Heat and Fluid Flow, 104: 0 109232, 2023. ISSN 0142-727X

    work page 2023

  7. [7]

    O. Hennigh. Lat-net: Compressing lattice Boltzmann flow simulations using deep neural networks . arXiv preprint arXiv:1705.09036, 2017

    work page Pith review arXiv 2017

  8. [8]

    Hsieh, S

    J.-T. Hsieh, S. Zhao, S. Eismann, L. Mirabella, and S. Ermon. Learning Neural PDE Solvers with Convergence Guarantees , 2019

    work page 2019

Show all 23 references
  1. [9]

    C. M. Jiang, S. Esmaeilzadeh, K. Azizzadenesheli, K. Kashinath, M. Mustafa, H. A. Tchelepi, P. Marcus, A. Anandkumar, et al. MeshfreeFlowNet: A Physics-Constrained Deep Continuous Space-Time Super-Resolution Framework . arXiv preprint arXiv:2005.01463, 2020

    2005 arXiv

  2. [10]

    Kochkov, J

    D. Kochkov, J. A. Smith, A. Alieva, Q. Wang, M. P. Brenner, and S. Hoyer. Machine learning--accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118 0 (21): 0 e2101784118, 2021

    2021

  3. [11]

    Z. Li, N. Kovachki, C. Choy, B. Li, J. Kossaifi, S. Otta, M. A. Nabian, M. Stadler, C. Hundt, K. Azizzadenesheli, and A. Anandkumar. Geometry-informed neural operator for large-scale 3d pdes. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Adva...

    2023

  4. [12]

    L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis. DeepXDE: A deep learning library for solving differential equations . arXiv preprint arXiv:1907.04502, 2019

    1907 arXiv

  5. [13]

    Majchrzak, K

    M. Majchrzak, K. Marciniak-Lukasiak, and P. Lukasiak. A survey on the application of machine learning in turbulent flow simulations. Energies, 16 0 (4), 2023

    2023

  6. [14]

    M. A. Nabian and H. Meidani. A deep learning solution approach for high-dimensional random differential equations. Probabilistic Engineering Mechanics, 57: 0 14--25, 2019

    2019

  7. [15]

    Oldenburg, F

    J. Oldenburg, F. Borowski, A. \"O ner, K.-P. Schmitz, and M. Stiehm. Geometry aware physics informed neural network surrogate for solving navier--stokes equation (gapinn). Advanced Modeling and Simulation in Engineering Sciences, 9 0 (1): 0 8, 2022

    2022

  8. [16]

    Patel, V

    Y. Patel, V. Mons, O. Marquet, and G. Rigas. Turbulence model augmented physics-informed neural networks for mean-flow reconstruction. Phys. Rev. Fluids, 9: 0 034605, Mar 2024. doi:10.1103/PhysRevFluids.9.034605

    2024 doi

  9. [17]

    Pioch, J

    F. Pioch, J. H. Harmening, A. M. Müller, F.-J. Peitzmann, D. Schramm, and O. el Moctar. Turbulence Modeling for Physics-Informed Neural Networks: Comparison of Different RANS Models for the Backward-Facing Step Flow . Fluids, 8 0 (2), 2023. ISSN 2311-5521. doi:10.3390/fluids8020043

    2023 doi

  10. [18]

    Raissi, P

    M. Raissi, P. Perdikaris, and 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: 0 686--707, 2019

    2019

  11. [19]

    Vinuesa and S

    R. Vinuesa and S. L. Brunton. Enhancing computational fluid dynamics with machine learning. Nature Computational Science, 2 0 (6): 0 358--366, 2022

    2022

  12. [20]

    R. Wang, K. Kashinath, M. Mustafa, A. Albert, and R. Yu. Towards physics-informed deep learning for turbulent flow prediction. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1457--1466, 2020

    2020

  13. [21]

    Warey, S

    A. Warey, S. Kaushik, B. Khalighi, M. Cruse, and G. Venkatesan. Data-driven prediction of vehicle cabin thermal comfort: using machine learning and high-fidelity simulation results. International Journal of Heat and Mass Transfer, 148: 0 119083, 2020

    2020

  14. [22]

    Zhang, B

    T. Zhang, B. Dey, P. Kakkar, A. Dasgupta, and A. Chakraborty. Frequency-compensated PINNs for Fluid-dynamic Design Problems . NeurIPS Workshop on Machine Learning for Engineering Modeling, Simulation, and Design, arXiv:2011.01456, 2020

    2011 arXiv

  15. [23]

    Zhang, B

    T. Zhang, B. Dey, K. Veeraraghavan, H. Kulkarni, and A. Chakraborty. Demystifying the Data Need of ML-surrogates for CFD Simulations . AAAI Workshop on AI to Accelerate Science and Engineering, arXiv:2205.08355, 2022

    2022 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.