Pith. sign in

REVIEW 4 major objections 6 minor 70 references

Treating neural-network flow images as finite-volume grids and training against numerical flux imbalance improves aerodynamic force predictions, especially when training data is scarce.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 19:50 UTC pith:OJTKT3GK

load-bearing objection Solid method paper: HLLC flux on image pixels as a supervised regularizer improves CD, especially with sparse AOA data; F_GT is CFD-derived residual matching, not a free conservation prior. the 4 major comments →

arxiv 2607.06587 v1 pith:OJTKT3GK submitted 2026-07-05 cs.CV physics.flu-dyn

CoFINN: Conservation Flux Informed Neural Networks for Physics Problems Governed by Conservation Laws

Adnan Harun Do\u{g}an , Mert Deniz , Hande Alemdar , \"Ozg\"ur U\u{g}ra\c{s} Baran This is my paper
classification cs.CV physics.flu-dyn
keywords physics-informed neural networksfinite-volume methodHLLC Riemann solvercompressible flowairfoil aerodynamicsCNN surrogatesconservation lawsdrag prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard CNN surrogates for airfoil flow learn pixel-wise resemblance to CFD fields but can violate conservation of mass, momentum, and energy, which shows up as large errors in lift and drag. CoFINN reinterprets every output pixel as a control-volume cell, evaluates convective fluxes across pixel faces with an HLLC Riemann solver, and adds the resulting conservation residual as a training loss alongside ordinary field matching. On transonic airfoil cases (Mach 0.7, Reynolds 6 million) that include shocks and high angles of attack, this conservation-flux term reduces drag coefficient error by roughly 15 percent on average and by as much as about 34 percent at extreme angles, with the largest relative gains in limited-data regimes. Inference cost stays the same as a plain CNN; only training pays a modest overhead. The authors present the method as architecture-agnostic and extensible to any conservation-law system that can be discretized on a structured grid.

Core claim

Embedding finite-volume conservation—by treating network output pixels as cells and penalizing imbalance of HLLC numerical fluxes against a CFD-derived ground-truth residual—produces flow-field surrogates whose integrated forces (especially drag) are substantially more accurate than pure data-driven training, particularly under data scarcity and at extreme angles of attack.

What carries the argument

CoFINN loss: CNN outputs are read as a structured mesh of finite-volume cells; first-order left/right states at pixel faces feed an HLLC Riemann solver; the net convective flux is compared to a precomputed ground-truth residual that already folds in viscous and wall-source terms, and the resulting MAE is mixed with ordinary field MAE via a weight λ.

Load-bearing premise

The claim rests on the idea that a supervised flux residual taken from CFD, plus first-order HLLC fluxes on a coarse 256-by-256 image grid that cannot resolve the boundary layer, is still a trustworthy regularizer for the integrated forces that matter.

What would settle it

Train identical architectures with and without the CoFINN term on the same held-out airfoil families at extreme angles of attack; if the far-field control-volume drag and lift errors do not improve (or if pure λ=1 training can be made to converge without data), the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • CNN, FNO, ViT and similar image-to-image flow surrogates can be made more force-accurate by adding a discrete conservation residual without changing their inference speed.
  • In data-limited aerodynamic design loops the conservation term acts as a physical regularizer that reduces reliance on dense multi-angle training sets.
  • Force evaluation via far-field momentum balance becomes more reliable even when near-wall pixels remain poorly resolved.
  • The same pixel-as-cell idea can be ported to other conservation-law systems (heat, mass transfer, hyperbolic systems) once an appropriate numerical flux is substituted.
  • Architecture choice becomes secondary to whether the loss landscape is shaped by discrete conservation.

Where Pith is reading between the lines

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

  • Because pure flux loss diverges, future work will likely need adaptive or curriculum weighting of λ rather than a fixed mix, or a self-supervised residual that does not rely on precomputed F_GT.
  • Graph or mesh-aware networks that keep true near-wall spacing could turn the same flux idea into a viscous-aware regularizer instead of an essentially inviscid one.
  • The method’s success on shocks suggests it may transfer to other discontinuous hyperbolic problems (traffic, shallow water, multiphase) where pixel-wise losses alone smear fronts.
  • If the far-field force integral is already the quantity of interest, one could supervise only the outer control-volume residual and drop full-field MAE for some design tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes CoFINN, a training framework that treats CNN (or other image-to-image) flow-field outputs as structured finite-volume grids and adds a loss that matches HLLC convective flux integrals on pixel faces to a CFD-precomputed residual F_GT (viscous fluxes plus solid-boundary sources). The method is evaluated on transonic airfoil RANS data (M=0.7, Re=6e6) with held-out airfoil families, multiple AOA training ranges, and four architectures (CNNFoil, FNO, ViT, Diffusion). The central claim is that this conservation-flux loss improves aerodynamic force prediction—especially drag—relative to pure MAE, with larger relative gains in limited-data regimes, while remaining architecture-agnostic and inference-efficient.

Significance. If the force gains are real and attributable to discrete conservation structure rather than extra CFD-derived residual supervision, the work is a useful bridge between modern FV/CFD practice and neural surrogates, and is more CFD-aligned than collocation PINNs. Strengths include multi-architecture evaluation, family-held-out test airfoils, AOA-range ablations, absolute CD tables, far-field force integration, and explicit acknowledgment that pure flux loss (λ=1) diverges and that F_GT is supervised. Those elements make the contribution potentially valuable for aerodynamic surrogate design, but the significance depends on clarifying what physical content the loss actually injects and on reporting that is not inflated by best-of-many selection.

major comments (4)
  1. §5, Eqs. (14)–(16) and (19): F_GT is precomputed from CFD as viscous fluxes plus B_internal (including wall forces that determine lift/drag). Matching F_pred to this fixed residual is supervised residual matching of the CFD momentum budget, not an unsupervised conservation prior. The paper notes this, but the abstract and Introduction still frame CoFINN as embedding conservation physics that acts as a physical regularizer distinct from data similarity. A control that isolates HLLC/FV structure (e.g., matching a non-physical residual, shuffled F_GT, or a simple force-coefficient head) is needed; without it, headline CD gains may be force-matching supervision rather than conservation structure per se.
  2. §7.3–7.6 and Tables 4–5: Every reported cell is the best of many training configurations (scheduler/EMA/Lookahead/SWA/early-stop/aug) × 3 seeds × λ grid, selected independently per (model, range, loss). This best-of-best protocol systematically inflates CoFINN vs MAE-only gaps and makes the abstract’s “~15% average / up to 34%” claims hard to interpret. Report mean±std over seeds for a fixed training recipe (or nested selection on validation only), and state the selection protocol in the main text.
  3. Abstract and §8 vs Tables 5–6: Headline numbers (“up to 34% at extreme AOA,” “~15% on average”) do not match the tabulated relative CD reductions (e.g., CNNFoil Full 53.8%→14.8%; Single 95.0%→13.8%). §8 also states that “Lift predictions exhibited even more consistent improvement across all training regimes,” which is contradicted by Table 5 (FNO/ViT lift often degrades on Narrow–Full; Diffusion worsens on Broad/Full). Align abstract/discussion claims with the tables and caveate architecture- and metric-dependence.
  4. §5.1 and force evaluation: First-order pixel-face HLLC on a 256×256 anisotropically stretched image grid omits resolved boundary-layer viscous physics; pure λ=1 diverges. The far-field CV force metric is sensible given near-wall limits, but then the claim that the method “significantly improv[es] physical consistency and conservation behavior” should be restricted to inviscid momentum transport / integrated forces, not general conservation fidelity. State this scope limit in the abstract and conclusions.
minor comments (6)
  1. Related Work largely duplicates the Introduction survey; compress one of the two.
  2. Eq. (2) text: “F_v is the viscous (convective) flux vector” is a typo; should be viscous/diffusive.
  3. Figure 3 labels F_y=−L and F_z=−D while the problem is 2D (x–y); clarify coordinate convention.
  4. Face areas A_east,west=1/255, A_north,south=2/255 are free parameters of the image mesh; justify or sensitivity-test them briefly.
  5. Diffusion results are mostly deferred to the supplement yet appear in main tables with large regressions under CoFINN on Broad/Full; either move a short main-text discussion or qualify “architecture-agnostic” more carefully.
  6. Filters |C_D|>0.01 and |C_L|>0.05 for relative errors should be stated in the abstract/metrics summary so relative % errors are not over-read near zero lift.

Circularity Check

0 steps flagged

No load-bearing circularity: held-out airfoil-family tests and CFD force benchmarks are independent of the training residual; F_GT is an honest supervised regularizer, not a tautological prediction.

full rationale

The paper’s central empirical claim—that adding a cell-wise HLLC flux residual loss improves CD/CL on transonic airfoil surrogates, especially in limited-AOA regimes—is evaluated on ten held-out airfoil families never seen in training (eval partition) and on a random val split, with forces compared to independent RANS CFD. That is a standard supervised generalization test, not a closed loop. The construction in §5 does make L_flux supervised: F_GT is precomputed from the same CFD labels as L_data (viscous fluxes plus solid-boundary sources), and the authors explicitly state that pure flux loss (λ=1) diverges and that CoFINN does not reduce data requirements. That is a framing/physics-content issue (how much of the gain is residual matching vs free-standing conservation), not circularity of derivation: at test time F_GT is unused; the network maps only the distance-field input to primitives, and forces are integrated from the predicted field. Self-citations to CNNFoil (Duru et al., overlapping authors) supply the baseline architecture, not a uniqueness theorem that forces the result. No fitted scalar is renamed as a prediction; λ is ablated. No step reduces the reported test-set force errors to the training inputs by construction. Score 1 reflects only the mild, disclosed dependence of the regularizer on CFD residuals—not a circular claim.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard compressible conservation laws and ideal-gas/RANS modeling from CFD practice, plus paper-specific modeling choices: image pixels as first-order FV cells, supervised F_GT residual targets from CFD, HLLC convective-only training fluxes, fixed λ mixing, and far-field force integration. Free parameters are mainly the loss weight λ and extensive training-configuration selection. CoFINN as a named loss framework is the main invented construct; it has no independent evidence outside these experiments.

free parameters (4)
  • λ (data vs flux loss weight) = typically 0.1–0.7 (best-of grid)
    Hand-chosen hyperparameter grid {0,0.1,0.3,0.5,0.7,1.0}; best λ selected per model/range; λ=1 diverges. Central reported gains depend on picking favorable λ.
  • Training configuration selector (scheduler/EMA/Lookahead/SWA/early-stop/aug × seed) = best of 6 (or 3 for Diffusion) × 3 seeds
    Each reported cell takes the best of multiple optimization recipes and 3 seeds; this is free selection that the headline errors depend on.
  • Pixel face areas Ae/w=1/255, An/s=2/255 = 1/255 and 2/255
    Ad hoc scaling from anisotropic 256×256 extraction domain; enters flux balance and force integrals.
  • Force-error filters |CD|>0.01, |CL|>0.05 = 0.01 / 0.05 thresholds
    Post-hoc exclusion of near-zero coefficients when averaging percentage errors; affects reported means.
axioms (5)
  • domain assumption Integral conservation laws for mass/momentum/energy with convective+viscous fluxes and ideal-gas EOS govern the target flows.
    §3 Eqs. (1)–(10); standard compressible CFD premise.
  • domain assumption Steady RANS CFD with Spalart–Allmaras at fixed M=0.7, Re=6e6 is an adequate ground truth for training and force evaluation.
    Dataset §7.1; all labels and F_GT come from this single regime.
  • ad hoc to paper Each CNN output pixel may be treated as a first-order finite-volume cell with face states equal to cell centers for HLLC.
    §4–5.1; enables differentiable flux loss on images but is a coarse discretization choice.
  • ad hoc to paper F_GT from CFD (viscous fluxes + B_internal) is a valid supervised target so that matching convective flux integrals enforces useful conservation for forces.
    §5 Eqs. (14)–(16); authors note this is supervised, not PINN-style self-supervision.
  • domain assumption Far-field rectangular CV momentum flux with negligible outer viscous flux equals aerodynamic force on the airfoil.
    §6 Eqs. (21)–(23); standard control-volume force idea under stated approximations.
invented entities (1)
  • CoFINN flux loss (pixel-grid HLLC conservation residual vs F_GT) no independent evidence
    purpose: Regularize network outputs toward discrete conservation and better integrated forces.
    Named framework and loss construction introduced here; evidence is only the reported airfoil experiments.

pith-pipeline@v1.1.0-grok45 · 33198 in / 3816 out tokens · 40832 ms · 2026-07-11T19:50:36.383652+00:00 · methodology

0 comments
read the original abstract

We present CoFINN (Conservation Flux Informed Neural Networks), a physics-informed deep learning framework for predicting compressible flow fields governed by conservation laws. Unlike conventional data-driven convolutional neural networks (CNNs), which optimize only pixel-wise similarity metrics, CoFINN embeds finite-volume conservation physics directly into the training process. Unlike classical physics-informed methods which enforce differential-equation residuals at collocation points through automatic differentiation, CoFINN adopts a finite-volume perspective consistent with modern CFD methodology. CoFINN interprets CNN output fields as structured computational grids, where each pixel represents a finite-volume cell, and enforces conservation consistency through sophisticated numerical flux calculations. The framework is evaluated on transonic flow prediction around airfoils at (M=0.7, Re=6 * 10^6), including challenging conditions involving shock waves and high angles of attack. Results show that CoFINN improves aerodynamic force prediction accuracy, reducing drag prediction error by up to 34% at extreme angles of attack and by approximately 15% on average across the test set. Improvements are particularly significant in limited-data regimes, demonstrating that the conservation-based loss acts as an effective physical regularizer. The proposed approach maintains the computational efficiency advantages of CNN surrogates while significantly improving physical consistency and conservation behavior. The framework is architecture-agnostic and extensible to broader classes of conservation-law-governed physical systems.

Figures

Figures reproduced from arXiv: 2607.06587 by Adnan Harun Do\u{g}an, Hande Alemdar, Mert Deniz, \"Ozg\"ur U\u{g}ra\c{s} Baran.

Figure 1
Figure 1. Figure 1: Pixels as control volumes concept fluid flow are the conservation laws: conservation of mass, momentum, and energy. For compress￾ible flows, an equation of state, such as the ideal gas law used in our case, is also essential. These conservation laws state that the total amount of a conserved property within a defined volume re￾mains constant over time, unless there is a net flow (flux) of that property acr… view at source ↗
Figure 2
Figure 2. Figure 2: Control volume cell and its neighbors UR, along with the face normal vector (n), to calculate the final convective flux Fc. This approach effectively accounts for the possibility of shock waves and contact discontinuities present in tran￾sonic flow fields: Fc = F(UL, UR, n) (11) 5 Conservation Flux Informed Neural Networks, CoFINN To embed physical principles into our CNN training, we utilize the integral … view at source ↗
Figure 3
Figure 3. Figure 3: Rectangular control volume around an airfoil used for force calculations [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Flow-field comparison for NACA 2421 at α = 0◦ (CNNFoil, Single training range). Columns: ground truth, MAE-only prediction, MAE-only error (signed), MAE+CoFINN prediction, MAE+CoFINN error (signed). Rows: pressure, u-velocity, v-velocity, temperature. Error colorbars are symmetric and shared between the two methods to enable direct visual comparison. At α = 0◦ , both models successfully predicts general ma… view at source ↗
Figure 5
Figure 5. Figure 5: Flow-field comparison for NACA 2421 at α = 12◦ (CNNFoil, Full training range). Same layout as [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: illustrates the predicted drag coefficient (CD) across the full range of angles of attack for the NACA 2421 airfoil. The ground truth data demonstrates the expected parabolic drag polar, with a sharp increase in drag at higher angles (α > 10◦ ) due to the onset of flow separation and massive wake formation.The baseline MAE-only model (λ = 0) struggles significantly across the domain. Most notably, it compl… view at source ↗
Figure 7
Figure 7. Figure 7: CL vs. angle of attack for NACA 2421 (eval partition, Full range, CNNFoil). Black: CFD ground truth, blue: MAE-only baseline (λ=0), red: CoFINN (best λ>0). The baseline misses negative-angle behavior while CoFINN tracks the CFD curve more closely. We additionally evaluate a conditional diffusion model based on the Bastek et al. Bastek et al. [2025] U-Net architecture. The implementation details and brief d… view at source ↗

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

70 extracted references · 26 canonical work pages

  1. [1]

    Parametric airfoil catalog, Part II: G

    Melin, Tomas , year=. Parametric airfoil catalog, Part II: G

  2. [2]

    2017 , publisher=

    Introduction to aircraft design , author=. 2017 , publisher=

    2017

  3. [3]

    2012 , publisher=

    Theory of wing sections: including a summary of airfoil data , author=. 2012 , publisher=

    2012

  4. [4]

    Progress in Aerospace Sciences , year=

    Neural networks: Applications and opportunities in aeronautics , author=. Progress in Aerospace Sciences , year=

  5. [5]

    Scientific Reports , year=

    Learning aerodynamics with neural network , author=. Scientific Reports , year=

  6. [6]

    , year =

    Rai, Man Mohan and Madavan, Nateri K. , year =. Aerodynamic Design Using Neural Networks , volume =. AIAA Journal , publisher =. doi:10.2514/2.938 , number =

    work page doi:10.2514/2.938

  7. [7]

    Independent Component Analyses, Wavelets, and Neural Networks , volume=

    Training data requirement for a neural network to predict aerodynamic coefficients , author=. Independent Component Analyses, Wavelets, and Neural Networks , volume=. 2003 , publisher=

    2003

  8. [8]

    Aerospace Science and Technology , year=

    Lift coefficient prediction at high angle of attack using recurrent neural network , author=. Aerospace Science and Technology , year=

  9. [9]

    2018 , journal =

    Jin, Xiaowei and Cheng, Peng and Chen, Wen-Li and Li, Hui , title =. 2018 , journal =. doi:10.1063/1.5024595 , type =

    work page doi:10.1063/1.5024595 2018

  10. [10]

    Super-resolution reconstruction of turbulent flows with machine learning , volume=

    Fukami, Kai and Fukagata, Koji and Taira, Kunihiko , year=. Super-resolution reconstruction of turbulent flows with machine learning , volume=. doi:10.1017/jfm.2019.238 , journal=

    doi:10.1017/jfm.2019.238 2019

  11. [11]

    A supervised neural network for drag prediction of arbitrary 2D shapes in laminar flows at low Reynolds number , journal =

    Jonathan Viquerat and Elie Hachem , keywords =. A supervised neural network for drag prediction of arbitrary 2D shapes in laminar flows at low Reynolds number , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.compfluid.2020.104645 , url =

    doi:10.1016/j.compfluid.2020.104645 2020

  12. [12]

    Journal of Fluid Mechanics , volume=

    Data-driven prediction of unsteady flow over a circular cylinder using deep learning , author=. Journal of Fluid Mechanics , volume=. 2019 , publisher=

    2019

  13. [13]

    arXiv preprint arXiv:1910.13532 , year=

    U-net architectures for fast prediction of incompressible laminar flows , author=. arXiv preprint arXiv:1910.13532 , year=

    Pith/arXiv arXiv 1910

  14. [14]

    Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    Guo, Xiaoxiao and Li, Wei and Iorio, Francesco , title =. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =. 2016 , doi =

    2016

  15. [15]

    2017 , journal =

    Singh, Anand Pratap and Medida, Shivaji and Duraisamy, Karthik , title =. 2017 , journal =. doi:10.2514/1.J055595 , type =

    doi:10.2514/1.j055595 2017

  16. [16]

    18th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference , year=

    A Convolutional Neural Network Approach to Training Predictors for Airfoil Performance , author=. 18th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference , year=. doi:10.2514/6.2017-3660 , publisher=

    work page doi:10.2514/6.2017-3660 2017

  17. [17]

    Prediction of aerodynamic flow fields using convolutional neural networks , volume =

    Bhatnagar, Saakaar and Afshar, Yaser and Pan, Shaowu and Duraisamy, Karthik and Kaushik, Shailendra , year =. Prediction of aerodynamic flow fields using convolutional neural networks , volume =. Computational Mechanics , publisher =. doi:10.1007/s00466-019-01740-0 , number =

    doi:10.1007/s00466-019-01740-0

  18. [18]

    Fast flow field prediction over airfoils using deep learning approach , volume =

    Sekar, Vinothkumar and Jiang, Qinghua and Shu, Chang and Khoo, Boo Cheong , year =. Fast flow field prediction over airfoils using deep learning approach , volume =. Physics of Fluids , publisher =. doi:10.1063/1.5094943 , number =

    work page doi:10.1063/1.5094943

  19. [19]

    Deep Learning Methods for Reynolds-Averaged

    Thuerey, Nils and Wei. Deep Learning Methods for Reynolds-Averaged. AIAA Journal , publisher =. 2020 , month = jan, pages =. doi:10.2514/1.j058291 , number =

    doi:10.2514/1.j058291 2020

  20. [20]

    Flow field prediction of supercritical airfoils via variational autoencoder based deep learning framework , volume =

    Wang, Jing and He, Cheng and Li, Runze and Chen, Haixin and Zhai, Chen and Zhang, Miao , year =. Flow field prediction of supercritical airfoils via variational autoencoder based deep learning framework , volume =. Physics of Fluids , publisher =. doi:10.1063/5.0053979 , number =

    work page doi:10.1063/5.0053979

  21. [21]

    CNNFOIL: convolutional encoder decoder modeling for pressure fields around airfoils , volume =

    Duru, Cihat and Alemdar, Hande and Baran, \". CNNFOIL: convolutional encoder decoder modeling for pressure fields around airfoils , volume =. Neural Computing and Applications , publisher =. 2020 , month = nov, pages =. doi:10.1007/s00521-020-05461-x , number =

    work page doi:10.1007/s00521-020-05461-x 2020

  22. [22]

    A generative deep learning framework for airfoil flow field prediction with sparse data , volume =

    WU, Haizhou and LIU, Xuejun and AN, Wei and LYU, Hongqiang , year =. A generative deep learning framework for airfoil flow field prediction with sparse data , volume =. Chinese Journal of Aeronautics , publisher =. doi:10.1016/j.cja.2021.02.012 , number =

    work page doi:10.1016/j.cja.2021.02.012 2021

  23. [23]

    The Aeronautical Journal (1968) , year=

    The development of a Riemann solver for the steady supersonic Euler equations , author=. The Aeronautical Journal (1968) , year=

    1968

  24. [24]

    A cost-effective CNN-BEM coupling framework for design optimization of horizontal axis tidal turbine blades

    Xu, Jian and Wang, Longyan and Yuan, Jianping and Shi, Jiali and Wang, Zilu and Zhang, Bowen and Luo, Zhaohui and Tan, Andy C C. A cost-effective CNN-BEM coupling framework for design optimization of horizontal axis tidal turbine blades. Energy (Oxf.)

  25. [25]

    Fast sparse flow field prediction around airfoils via multi-head perceptron based deep learning architecture

    Zuo, Kuijun and Bu, Shuhui and Zhang, Weiwei and Hu, Jiawei and Ye, Zhengyin and Yuan, Xianxu. Fast sparse flow field prediction around airfoils via multi-head perceptron based deep learning architecture. Aerosp. Sci. Technol

  26. [26]

    A comparative study of learning techniques for the compressible aerodynamics over a transonic RAE2822 airfoil

    Catalani, Giovanni and Costero, Daniel and Bauerheim, Michael and Zampieri, Luca and Chapin, Vincent and Gourdain, Nicolas and Baqu \'e , Pierre. A comparative study of learning techniques for the compressible aerodynamics over a transonic RAE2822 airfoil. Comput. Fluids

  27. [27]

    Machine learning-based surrogate modeling approaches for fixed-wing store separation

    Peters, Nicholas and Wissink, Andrew and Ekaterinaris, John. Machine learning-based surrogate modeling approaches for fixed-wing store separation. Aerosp. Sci. Technol

  28. [28]

    arXiv , author =

    U-. arXiv , author =. 2022 , note =. doi:10.48550/arXiv.2109.03697 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.03697 2022

  29. [29]

    arXiv preprint arXiv:2404.07200 , author =

    Toward a. arXiv preprint arXiv:2404.07200 , author =. 2024 , file =

    Pith/arXiv arXiv 2024

  30. [30]

    , year =

    Bastek, Jan-Hendrik and Sun, WaiChing and Kochmann, Dennis M. , year =. Physics-Informed Diffusion Models , journal =

  31. [31]

    Communications in Computational Physics , author =

    Adaptive. Communications in Computational Physics , author =. 2025 , doi =

    2025

  32. [32]

    arXiv.org , author =

    Generative prediction of flow fields around an obstacle using the diffusion model , url =. arXiv.org , author =. 2024 , file =

    2024

  33. [33]

    Neural fields for rapid aircraft aerodynamics simulations , volume =

    Catalani, Giovanni and Agarwal, Siddhant and Bertrand, Xavier and Tost, Frederic and Bauerheim, Michael and Morlier, Joseph , year =. Neural fields for rapid aircraft aerodynamics simulations , volume =. doi:10.1038/s41598-024-76983-w , journal =

    work page doi:10.1038/s41598-024-76983-w

  34. [34]

    arXiv preprint arXiv:2411.17164 , author =

    X-. arXiv preprint arXiv:2411.17164 , author =. 2024 , file =

    Pith/arXiv arXiv 2024

  35. [35]

    Computers & Fluids , author =

    A deep learning approach for the transonic flow field predictions around airfoils , volume =. Computers & Fluids , author =. 2022 , keywords =. doi:10.1016/j.compfluid.2022.105312 , urldate =

    work page doi:10.1016/j.compfluid.2022.105312 2022

  36. [36]

    Physics of Fluids , author =

    A deep learning based prediction approach for the supercritical airfoil at transonic speeds , volume =. Physics of Fluids , author =. 2021 , pages =. doi:10.1063/5.0060604 , number =

    work page doi:10.1063/5.0060604 2021

  37. [37]

    Physics-informed neural networks for solving moving interface flow problems using the level set approach , url =

    Mullins, Mathieu and Kamil, Hamza and Fahsi, Adil and Soulaimani, Azzeddine , month = feb, journal=. Physics-informed neural networks for solving moving interface flow problems using the level set approach , url =. 2025 , publisher=

    2025

  38. [38]

    2025 , file =

    arXiv preprint arXiv:2503.16323 , author =. 2025 , file =

    Pith/arXiv arXiv 2025

  39. [39]

    International Mechanical Engineering Congress and Exposition-India , volume=

    Accelerated Airfoil Design Using Neural Network Approaches , author=. International Mechanical Engineering Congress and Exposition-India , volume=. 2025 , organization=

    2025

  40. [40]

    arXiv preprint arXiv:2508.01463 , author =

    Extended. arXiv preprint arXiv:2508.01463 , author =. 2025 , file =

    Pith/arXiv arXiv 2025

  41. [41]

    Computers & Fluids , author =

    A general deep transfer learning framework for predicting the flow field of airfoils with small data , volume =. Computers & Fluids , author =. 2023 , keywords =. doi:10.1016/j.compfluid.2022.105738 , urldate =

    work page doi:10.1016/j.compfluid.2022.105738 2023

  42. [42]

    Heliyon , author =

    Airfoil aerodynamic performance prediction using machine learning and surrogate modeling , volume =. Heliyon , author =. 2024 , keywords =. doi:10.1016/j.heliyon.2024.e29377 , number =

    work page doi:10.1016/j.heliyon.2024.e29377 2024

  43. [43]

    Fluids , author =

    Deep. Fluids , author =. 2022 , note =. doi:10.3390/fluids7110344 , language =

    work page doi:10.3390/fluids7110344 2022

  44. [44]

    Journal of Computational Physics , author =

    Deep learning of free boundary and. Journal of Computational Physics , author =. 2021 , keywords =. doi:10.1016/j.jcp.2020.109914 , urldate =

    work page doi:10.1016/j.jcp.2020.109914 2021

  45. [45]

    Physics of Fluids , author =

    Fast aerodynamics prediction of laminar airfoils based on deep attention network , volume =. Physics of Fluids , author =. 2023 , pages =. doi:10.1063/5.0140545 , number =

    work page doi:10.1063/5.0140545 2023

  46. [46]

    Applied Sciences , author =

    Fast. Applied Sciences , author =. 2022 , note =. doi:10.3390/app122312075 , language =

    work page doi:10.3390/app122312075 2022

  47. [47]

    Aerospace Science and Technology , author =

    Fast pressure distribution prediction of airfoils using deep learning , volume =. Aerospace Science and Technology , author =. 2020 , keywords =. doi:10.1016/j.ast.2020.105949 , urldate =

    work page doi:10.1016/j.ast.2020.105949 2020

  48. [48]

    AIAA Journal , author =

    Flight. AIAA Journal , author =. 2024 , note =. doi:10.2514/1.J063992 , number =

    work page doi:10.2514/1.j063992 2024

  49. [49]

    Entropy , author =

    Improved. Entropy , author =. 2023 , note =. doi:10.3390/e25040675 , language =

    work page doi:10.3390/e25040675 2023

  50. [50]

    Applied Sciences , author =

    Learning. Applied Sciences , author =. 2024 , note =. doi:10.3390/app142210685 , language =

    work page doi:10.3390/app142210685 2024

  51. [51]

    2024 , author =

    Airfoil. 2024 , author =

    2024

  52. [52]

    Computer Methods in Applied Mechanics and Engineering , author =

    Physics-. Computer Methods in Applied Mechanics and Engineering , author =. 2025 , keywords =. doi:10.1016/j.cma.2025.118015 , urldate =

    work page doi:10.1016/j.cma.2025.118015 2025

  53. [53]

    Physics of Fluids , author =

    Physics-informed neural network compression mechanism for airfoil flow field prediction , volume =. Physics of Fluids , author =. 2025 , pages =. doi:10.1063/5.0255692 , number =

    work page doi:10.1063/5.0255692 2025

  54. [54]

    Physics of Fluids , author =

    Physics-informed neural networks for fully non-linear free surface wave propagation , volume =. Physics of Fluids , author =. 2024 , pages =. doi:10.1063/5.0210591 , number =

    work page doi:10.1063/5.0210591 2024

  55. [55]

    Physics of Fluids , author =

    Physics-informed neural networks for incompressible flows with moving boundaries , volume =. Physics of Fluids , author =. 2024 , pages =. doi:10.1063/5.0186809 , number =

    work page doi:10.1063/5.0186809 2024

  56. [56]

    Computers & Fluids , author =

    Towards high-accuracy deep learning inference of compressible flows over aerofoils , volume =. Computers & Fluids , author =. 2023 , keywords =. doi:10.1016/j.compfluid.2022.105707 , urldate =

    work page doi:10.1016/j.compfluid.2022.105707 2023

  57. [57]

    International Journal for Numerical Methods in Fluids , author =

    Deep. International Journal for Numerical Methods in Fluids , author =. 2025 , note =. doi:10.1002/fld.5375 , language =

    work page doi:10.1002/fld.5375 2025

  58. [58]

    SIAM Journal on Scientific Computing , author =

    Physics-. SIAM Journal on Scientific Computing , author =. 2023 , note =. doi:10.1137/22M1517081 , number =

    doi:10.1137/22m1517081 2023

  59. [59]

    Energies , volume =

    A Review of Physics-Informed Machine Learning in Fluid Mechanics , author =. Energies , volume =. 2023 , month = feb, doi =

    2023

  60. [60]

    30th Aerospace Sciences Meeting and Exhibit , pages=

    A one-equation turbulence model for aerodynamic flows , author=. 30th Aerospace Sciences Meeting and Exhibit , pages=. 1992 , publisher=

    1992

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Lookahead Optimizer: k steps forward, 1 step back , author=. Advances in Neural Information Processing Systems , volume=. 2019 , publisher=

    2019

  62. [62]

    and di Bernardo, Mario and Russo, Lucia and Siettos, Constantinos I

    Patsatzis, Dimitrios G. and di Bernardo, Mario and Russo, Lucia and Siettos, Constantinos I. , journal=. 2025 , doi=

    2025

  63. [63]

    Physics of Fluids , volume=

    Physics-informed neural networks for inviscid transonic flows around an airfoil , author=. Physics of Fluids , volume=. 2025 , doi=

    2025

  64. [64]

    Physics of Fluids , volume=

    Deep learning-based predictive modeling of transonic flow over an airfoil , author=. Physics of Fluids , volume=. 2024 , doi=

    2024

  65. [65]

    and Spruce, Michael and Speares, William , journal=

    Toro, Eleuterio F. and Spruce, Michael and Speares, William , journal=. Restoration of the contact surface in the. 1994 , doi=

    1994

  66. [66]

    International Conference on Learning Representations (ICLR) , year=

    Fourier neural operator for parametric partial differential equations , author=. International Conference on Learning Representations (ICLR) , year=

  67. [67]

    International Journal of Computational Fluid Dynamics , volume=

    Prediction of Flow Field Over Airfoils Based on Transformer Neural Network , author=. International Journal of Computational Fluid Dynamics , volume=. 2023 , publisher=

    2023

  68. [68]

    International Conference on Learning Representations (ICLR) , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations (ICLR) , year=

  69. [69]

    International Conference on Learning Representations , year=

    Flow matching for generative modeling , author=. International Conference on Learning Representations , year=

  70. [70]

    Low Reynolds Number Aerodynamics , year =

    Mark Drela , title =. Low Reynolds Number Aerodynamics , year =