pith. sign in

arxiv: 2605.07131 · v2 · pith:F6VZU2KVnew · submitted 2026-05-08 · ⚛️ physics.flu-dyn

A fast Physics-Informed Neural Networks based approach to the 2D design of turbine blades

Yuan Huang , Francesca di Mare This is my paper

Pith reviewed 2026-05-25 06:38 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords Physics-informed neural networksTurbomachinery bladesAerodynamic screeningEuler equationsBoundary condition relaxationDynamic loss weightingBlade design optimisation
0
0 comments X

The pith

A progressive PINN framework screens families of turbine blades with CFD-comparable accuracy at lower cost.

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

The paper introduces a progressive Euler-PINN method that gradually relaxes boundary conditions from empty tunnel flow to full outlet static pressure while applying geometry-aware dynamic loss weighting near curved surfaces. This single workflow handles ten NACA 6 blade variants across thirty subsonic operating points. A reader would care because conventional CFD remains too slow for wide preliminary design sweeps in turbomachinery for energy systems. The approach therefore positions itself as a practical surrogate for two-dimensional blade pre-design and optimisation.

Core claim

The progressive Euler-PINN framework, which combines gradual boundary-condition relaxation from tunnel flow without a blade to full outlet static pressure and a geometry-aware dynamic loss-weighting scheme that intensifies penalties near highly curved boundaries, achieves CFD-comparable accuracy for pressure and velocity fields while reducing the computational cost required for family-wide blade screening across multiple operating conditions.

What carries the argument

Progressive Euler-PINN framework that gradually relaxes boundary conditions and applies geometry-aware dynamic loss weighting near curved boundaries.

If this is right

  • One trained workflow can evaluate ten blade variants across thirty operating points.
  • Computational cost for family-wide screening drops below that of repeated CFD runs.
  • Pressure and velocity fields reach accuracy levels comparable to conventional CFD.
  • The method serves as a practical surrogate for two-dimensional turbomachinery pre-design.
  • The same single workflow covers multiple operating conditions without separate retraining.

Where Pith is reading between the lines

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

  • Similar progressive relaxation might stabilise PINNs in other external-flow problems with sharp geometry changes.
  • The loss-weighting scheme could be tested on three-dimensional blade passages to check whether the same principles hold without major reformulation.
  • Integration into an outer optimisation loop would allow automated generation of blade families that meet multiple off-design targets.

Load-bearing premise

The combination of progressive boundary-condition relaxation and geometry-aware dynamic loss weighting will produce stable convergence and engineering-grade accuracy for complex blade geometries and off-design conditions without case-by-case hyperparameter retuning.

What would settle it

A new blade geometry or off-design point where the trained network either fails to converge or produces pressure and velocity fields that deviate substantially from CFD results without any hyperparameter adjustment.

Figures

Figures reproduced from arXiv: 2605.07131 by Francesca di Mare, Yuan Huang.

Figure 1
Figure 1. Figure 1: Initial collocation layout for NACA 65-1010. Orange dots are interior Sobol [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Main chain with special branch: 0610 to 1010 via 5 [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Spatial distribution of the geometry-aware weight field (case 65-1010). Hottest [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Velocity contours for the three training strategies. Cold start is non-physical; [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Airfoil 65-1010, α = 5◦ , pout = 0.80 bar. Rows: PINN prediction, SharC CFD, relative error. Columns: density ρ, Mach number M, static pressure p. 16 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Surface pressure coefficient Cp for NACA65-1010 (α = 0◦ , pout = 0.8bar). PINN (solid blue) and CFD (dashed black) for pressure side (positive Cp) and suction side (negative Cp). large errors appear only near the leading edge and trailing edge where |Cp| approaches zero. 17 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training loss. Every loss bump refer to a boundary condition restart. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: CL vs. design lift m/10. PINN points are connected per (α, pout); CFD is shown as “×” without lines. NACA1010 with AoA 0 doesn’t have a steady state Euler solution Observations.. Our results highlight a clear hierarchy in how different design￾and operating-variables influence the accuracy of the PINN predictions. First and most significantly, blade geometry — in particular thickness and cam￾ber/curvature —… view at source ↗
read the original abstract

Rapid aerodynamic screening of turbomachinery blades across wide operating envelopes remains a major computational bottleneck in preliminary design, particularly for energy-conversion and storage systems such as emerging Carnot batteries. Physics-informed neural networks (PINNs) offer a mesh-free alternative to conventional CFD, yet convergence and accuracy often deteriorate for complex blade geometries and off-design flows. We propose a progressive Euler-PINN framework that (i) gradually relaxes boundary conditions from tunnel flow without a blade to full outlet static pressure, and (ii) employs a geometry-aware dynamic loss-weighting scheme that intensifies residual penalties near highly curved boundaries. To the best of our knowledge, this is the first study to deploy a single PINN workflow for large-scale, engineering-grade screening of turbomachinery blade families across multiple operating conditions, covering ten NACA6 variants and 30 subsonic operating points. The proposed framework achieves CFD-comparable accuracy for pressure and velocity fields while reducing the computational cost required for family-wide blade screening. These results establish the method as a practical surrogate for two-dimensional turbomachinery blade pre-design and optimisation.

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 / 1 minor

Summary. The paper proposes a progressive Euler-PINN framework for 2D turbomachinery blade design. It combines gradual relaxation of boundary conditions (from tunnel flow without a blade to full outlet static pressure) with a geometry-aware dynamic loss-weighting scheme that intensifies penalties near high-curvature regions. The central claim is that this single workflow delivers CFD-comparable accuracy in pressure and velocity fields for ten NACA6 variants across 30 subsonic operating points while lowering the cost of family-wide screening.

Significance. If substantiated, the approach could offer a mesh-free surrogate for preliminary blade screening in applications such as Carnot batteries, where rapid evaluation over wide envelopes is needed. The emphasis on a unified workflow without per-geometry retuning would be a practical advantage over conventional CFD if the accuracy and cost claims are shown to hold uniformly.

major comments (2)
  1. [Abstract] Abstract: the claim that a single PINN workflow suffices for all ten NACA6 variants and 30 operating points rests on the unverified assumption that one fixed relaxation schedule and set of geometry-aware weighting coefficients produces stable, CFD-comparable solutions without per-case adjustment. No ablation or sensitivity study is described that holds these hyperparameters fixed while varying camber, thickness, or incidence; this directly undermines the 'single workflow' and family-wide cost-reduction assertions.
  2. [Method] The description of the dynamic loss-weighting scheme indicates intensification near high-curvature boundaries, yet no quantitative evidence (e.g., residual histories or error tables across the full matrix of geometries and conditions) is provided to confirm that the same weighting exponents remain effective without retuning. This is load-bearing for the engineering-grade accuracy claim.
minor comments (1)
  1. [Abstract] The abstract should explicitly state the quantitative error metrics (e.g., L2 norms, maximum errors) and reference CFD solver used to support the 'CFD-comparable accuracy' statement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and will incorporate revisions to strengthen the supporting evidence for our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that a single PINN workflow suffices for all ten NACA6 variants and 30 operating points rests on the unverified assumption that one fixed relaxation schedule and set of geometry-aware weighting coefficients produces stable, CFD-comparable solutions without per-case adjustment. No ablation or sensitivity study is described that holds these hyperparameters fixed while varying camber, thickness, or incidence; this directly undermines the 'single workflow' and family-wide cost-reduction assertions.

    Authors: We agree that an explicit ablation study holding the relaxation schedule and weighting coefficients fixed would provide stronger substantiation. The manuscript applies the identical progressive Euler-PINN framework and geometry-aware weighting to all ten NACA6 variants across 30 operating points and reports consistent CFD-comparable accuracy without per-case retuning. To address the concern directly, we will add a sensitivity analysis in the revised manuscript that varies camber, thickness, and incidence while keeping the hyperparameters fixed and tabulates the resulting accuracy metrics. revision: yes

  2. Referee: [Method] The description of the dynamic loss-weighting scheme indicates intensification near high-curvature boundaries, yet no quantitative evidence (e.g., residual histories or error tables across the full matrix of geometries and conditions) is provided to confirm that the same weighting exponents remain effective without retuning. This is load-bearing for the engineering-grade accuracy claim.

    Authors: We acknowledge that residual histories and error tables across the full matrix would offer clearer confirmation of the fixed weighting scheme. The current results section reports aggregate accuracy for the complete set of cases, but we will expand it in revision to include residual convergence histories and detailed error tables broken down by geometry and operating condition to demonstrate consistent performance without retuning. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is a forward simulation technique

full rationale

The paper introduces a progressive Euler-PINN framework using boundary-condition relaxation and geometry-aware dynamic loss weighting for blade flow prediction. No equations, fitted parameters, or self-citations are shown that reduce the claimed accuracy or cost savings to a definition, a renamed input, or a self-referential fit. The central results are presented as empirical outcomes on NACA6 cases rather than tautological consequences of the method definition. This is the expected finding for a mesh-free forward solver whose performance is validated externally against CFD.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated. The framework implicitly assumes that the Euler equations plus the chosen boundary conditions are sufficient to capture the target flows and that the neural-network optimizer will find a solution satisfying both physics and data.

pith-pipeline@v0.9.0 · 5721 in / 1107 out tokens · 19994 ms · 2026-05-25T06:38:40.538360+00:00 · methodology

Review history (2 revisions) →

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

29 extracted references · 29 canonical work pages

  1. [1]

    Machine learning in fluid dynamics—physics-informed neural net- works (pinns) using sparse data: A review.Fluids, 10(9):226, 2025

    Mjalled Ali, Philippe Miron, Martin Mönnigmann, Bukharin Nikolay, et al. Machine learning in fluid dynamics—physics-informed neural net- works (pinns) using sparse data: A review.Fluids, 10(9):226, 2025

    work page 2025

  2. [2]

    Botarelli et al

    T. Botarelli et al. Using physics-informed neural networks for solving navier–stokes problems in near real-world geometries.Comput. Methods Appl. Mech. Eng., 2025. online first / in-press (as of 2025)

    work page 2025

  3. [3]

    Physics-informed neural networks for fluid mechanics: a review.Acta Mechanica Sinica, 37(12):1727–1738, 2021

    Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks for fluid mechanics: a review.Acta Mechanica Sinica, 37(12):1727–1738, 2021

    work page 2021

  4. [4]

    A solver for subsonic flow around airfoils based on physics-informed neural networks and mesh transformation.Physics of Fluids, 36(2), 2024

    Wenbo Cao, Jiahao Song, and Weiwei Zhang. A solver for subsonic flow around airfoils based on physics-informed neural networks and mesh transformation.Physics of Fluids, 36(2), 2024

    work page 2024

  5. [5]

    Pi-Yueh Chuang and Lorena A. Barba. Experience report of physics- informed neural networks in fluid simulations: pitfalls and frustration. arXiv preprint, 2022

    work page 2022

  6. [6]

    J. K. Cornelius, T. Ågren, N. Peters, et al. Palmo: An overflow machine learning airfoil performance database. Technical report, NASA Ames Research Center, 2024

    work page 2024

  7. [7]

    Physics-informed neu- ral networks with adaptive localized artificial viscosity.Journal of Com- putational Physics, 489:112265, 2023

    Emilio Jose Rocha Coutinho, Marcelo Dall’Aqua, Levi McClenny, Ming Zhong, Ulisses Braga-Neto, and Eduardo Gildin. Physics-informed neu- ral networks with adaptive localized artificial viscosity.Journal of Com- putational Physics, 489:112265, 2023. 23

    work page 2023

  8. [8]

    Frate, A

    Olivier Dumont, G.F. Frate, A. Pillai, S. Lecompte, M. De Paepe, and V. Lemort. Carnot battery technology: A state-of-the-art review.Jour- nal of Energy Storage, 32:101756, 2020

    work page 2020

  9. [9]

    Physics-informed neural networks for solving reynolds-averaged navier–stokes equations.Physics of Fluids, 34(7):075117, 2022

    Hamidreza Eivazi, Mojtaba Tahani, Philipp Schlatter, and Ricardo Vin- uesa. Physics-informed neural networks for solving reynolds-averaged navier–stokes equations.Physics of Fluids, 34(7):075117, 2022

    work page 2022

  10. [10]

    Physics-informed neural network compression mechanism for airfoil flow field prediction.Physics of Fluids, 37(3), 2025

    Hongyu Huang, Yiyang Ye, Bohan Zhang, Zhijiang Xie, Fei Xu, and Chao Chen. Physics-informed neural network compression mechanism for airfoil flow field prediction.Physics of Fluids, 37(3), 2025

    work page 2025

  11. [11]

    Jagtap, Zhiping Mao, Nikolaus A

    Ameya D. Jagtap, Zhiping Mao, Nikolaus A. Adams, and George Em Karniadakis. Physics-informed neural networks for inverse problems in supersonic flows.Journal of Computational Physics, 466:111402, 2022. arXiv:2202.11821

    work page arXiv 2022

  12. [12]

    A surrogate-model based multidisciplinary shape optimization method withapplicationtoa2dsubsonicairfoil.Computers & Fluids, 36(3):520– 529, 2007

    J-C Jouhaud, Pierre Sagaut, Marc Montagnac, and Julien Laurenceau. A surrogate-model based multidisciplinary shape optimization method withapplicationtoa2dsubsonicairfoil.Computers & Fluids, 36(3):520– 529, 2007

    work page 2007

  13. [13]

    Characterizing possible failure modes in physics- informed neural networks.Advances in neural information processing systems, 34:26548–26560, 2021

    Aditi Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W Mahoney. Characterizing possible failure modes in physics- informed neural networks.Advances in neural information processing systems, 34:26548–26560, 2021

    work page 2021

  14. [14]

    Detached eddy simulation of a radial turbine operated with supercritical carbon dioxide.International Journal of Turbomachinery, Propulsion and Power, 10(4):43, 2025

    Benedikt Lea, Federico Lo Presti, Wojciech Sadowski, and Francesca di Mare. Detached eddy simulation of a radial turbine operated with supercritical carbon dioxide.International Journal of Turbomachinery, Propulsion and Power, 10(4):43, 2025

    work page 2025

  15. [15]

    Machine learning in aerodynamic shape optimization.Progress in Aerospace Sciences, 134:100849, 2022

    Jichao Li, Xiaosong Du, and Joaquim RRA Martins. Machine learning in aerodynamic shape optimization.Progress in Aerospace Sciences, 134:100849, 2022

    work page 2022

  16. [16]

    Key components for carnot battery: Technology review, technical barriers and selection criteria

    Ting Liang, Andrea Vecchi, Kai Knobloch, Adriano Sciacovelli, Kurt Engelbrecht, Yongliang Li, and Yulong Ding. Key components for carnot battery: Technology review, technical barriers and selection criteria. Renewable and Sustainable Energy Reviews, 163:112478, 2022. 24

    work page 2022

  17. [17]

    Quantification of airfoil geometry-induced aerodynamic uncertainties— comparison of approaches.SIAM/ASA Journal on Uncertainty Quan- tification, 5(1):334–352, 2017

    Dishi Liu, Alexander Litvinenko, Claudia Schillings, and Volker Schulz. Quantification of airfoil geometry-induced aerodynamic uncertainties— comparison of approaches.SIAM/ASA Journal on Uncertainty Quan- tification, 5(1):334–352, 2017

    work page 2017

  18. [18]

    A comprehensive review of advances in physics- informed neural networks and their applications in complex fluid dy- namics.Physics of Fluids, 36(10):101301, 2024

    Wenqiang Lou et al. A comprehensive review of advances in physics- informed neural networks and their applications in complex fluid dy- namics.Physics of Fluids, 36(10):101301, 2024

    work page 2024

  19. [19]

    Fundamental flaws of physics-informed neural networks and explainability methods in engineering systems.Computers & Industrial Engineering, page 111704, 2025

    MZ Naser. Fundamental flaws of physics-informed neural networks and explainability methods in engineering systems.Computers & Industrial Engineering, page 111704, 2025

    work page 2025

  20. [20]

    Geometry aware physics informed neural network sur- rogate for solving navier–stokes equation.Adv

    Jan Oldenburg, Finja Borowski, Alper Öner, Klaus-Peter Schmitz, and Michael Stiehm. Geometry aware physics informed neural network sur- rogate for solving navier–stokes equation.Adv. Model. Simul. Eng. Sci., 9(1):1–15, 2022

    work page 2022

  21. [21]

    Automatic differentiation in pytorch

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017

    work page 2017

  22. [22]

    Investigationof physics-informed neural networks based solution techniques for internal flows

    PascalPost, BenjaminWinhart, andFrancescaDiMare. Investigationof physics-informed neural networks based solution techniques for internal flows. InVolume 10C: Turbomachinery — Design Methods and CFD Modeling for Turbomachinery; Ducts, Noise, and Component Interac- tions, page V10CT32A011. American Society of Mechanical Engineers

    work page

  23. [23]

    Physics- informed neural networks: A deep learning framework for solving for- ward and inverse problems involving nonlinear partial differential equa- tions.J

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics- informed neural networks: A deep learning framework for solving for- ward and inverse problems involving nonlinear partial differential equa- tions.J. Comput. Phys., 378:686–707, 2019

    work page 2019

  24. [24]

    Challenges in training pinns: A loss landscape perspective

    Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in training pinns: A loss landscape perspective. 2024

    work page 2024

  25. [25]

    Ilya M. Sobol’. Distribution of points in a cube and approximate evalu- ation of integrals.USSR Computational Mathematics and Mathematical Physics, 7:86–112, 1967. 25

    work page 1967

  26. [26]

    Sukumar and A

    N. Sukumar and A. Srivastava. Exact imposition of boundary condi- tions with distance functions in physics-informed deep neural networks. Comput. Methods Appl. Mech. Eng., 388:114333, 2021

    work page 2021

  27. [27]

    R. C. Swanson and S. Langer. Comparison of naca 0012 laminar flow solutions: Structured and unstructured grid methods.NASA Technical Memorandum NASA/TM-2016-219003, 2016

    work page 2016

  28. [28]

    Naca 6-series airfoil profile: Nomencla- ture and generation rules

    NASA / OpenVSP Tutorial. Naca 6-series airfoil profile: Nomencla- ture and generation rules. Online tutorial, VSP ground school, 2024. Accessed: 2025-12-02

    work page 2024

  29. [29]

    A comprehensive review of advances in physics-informed neural networks and their applications in complex fluid dynamics.Physics of Fluids, 36(10):101301, 2024

    Chi Zhao, Feifei Zhang, Wenqiang Lou, Xi Wang, and Jianyong Yang. A comprehensive review of advances in physics-informed neural networks and their applications in complex fluid dynamics.Physics of Fluids, 36(10):101301, 2024. 26

    work page 2024