pith. sign in

arxiv: 2606.24265 · v1 · pith:TFZJBKATnew · submitted 2026-06-23 · 💻 cs.CE · cs.AI

Neural Network-Based Parametric Model Reduction for Predicting Turbulent Flow for Different Vehicle Geometries

Kazuto Ando , Rahul Bale , Akiyoshi Kuroda , Makoto Tsubokura This is my paper

Pith reviewed 2026-06-25 22:06 UTC · model grok-4.3

classification 💻 cs.CE cs.AI
keywords model reductionneural networksvariational autoencoderturbulent flowsvehicle aerodynamicsreduced-order modelinghigh Reynolds number
0
0 comments X

The pith

Adding a variational autoencoder to neural-network model reduction enables robust prediction of turbulent flows around different vehicle geometries.

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

The paper extends prior work on neural-network-based reduced-order models by incorporating a variational autoencoder. This extension is used to test robustness in simulations of high-Reynolds-number flows around several vehicle bodies that have different shapes. The evaluation centers on how accurately the model reconstructs the generation of vortices at various spatial and temporal scales, especially near the rear of the vehicles. A reader might care because running full simulations for every possible body design is expensive, so better reduction methods could speed up the design process. The framework retains its ability to train on large datasets using distributed computing.

Core claim

By incorporating a variational autoencoder, the neural-network-based model reduction can assess robustness in high-Reynolds-number flows around multiple vehicle bodies with varying geometries, evaluating the reconstruction accuracy of vortex generation across different spatial and temporal scales using a compact latent representation with focus on the flow near the rear end.

What carries the argument

Variational autoencoder that learns a compact latent representation of the flow field to enable nonlinear projection in the reduced-order model.

If this is right

  • The reduced model maintains accuracy for vortex reconstruction in turbulent conditions.
  • It applies to multiple vehicle geometries without retraining from scratch.
  • High-resolution data can be handled efficiently through the existing parallel training setup.
  • Robustness is demonstrated for flows at high Reynolds numbers.

Where Pith is reading between the lines

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

  • The approach might generalize to other design problems with varying parameters beyond vehicle shapes.
  • New vehicle designs could be evaluated by interpolating in the latent space without new full simulations.
  • Similar autoencoder enhancements could improve model reduction in other engineering simulations like heat transfer or structural analysis.

Load-bearing premise

The compact latent representation learned by the variational autoencoder captures the essential dynamics of vortex generation for different vehicle geometries without losing critical information at high Reynolds numbers.

What would settle it

A direct comparison showing that the reduced model cannot accurately reconstruct vortex patterns for an unseen vehicle geometry at high Reynolds number would falsify the robustness claim.

read the original abstract

Numerical simulations in industrial applications often require performing numerous high-precision computations parameterized by specific experimental conditions. For instance, in vehicle body design, aerodynamic simulations are essential for evaluating the aerodynamic characteristics of various proposed body geometries. However, computational resource constraints often become a bottleneck. Therefore, achieving the desired accuracy while minimizing computational cost is crucial. To address this challenge, model reduction methods have been developed to decrease the degrees of freedom by constraining the possible states of a physical system to a lower-dimensional subspace. In particular, reduction techniques that project the system onto a nonlinear subspace using neural networks have been actively studied. Our previous research developed a reduced-order model that integrates neural-network-based model reduction with a time-evolution method, implemented as a distributed parallel training framework to process high-resolution flow field data efficiently. In this study, we extend this reduction approach by incorporating a variational autoencoder to assess its robustness in high-Reynolds-number flows around multiple vehicle bodies with varying geometries. Specifically, we evaluate the reconstruction accuracy of vortex generation across different spatial and temporal scales using a compact latent representation, with a particular focus on the flow behavior near the rear end of the vehicle body.

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 manuscript extends prior neural-network-based reduced-order modeling by integrating a variational autoencoder (VAE) to create a parametric model for high-Reynolds-number turbulent flows around multiple vehicle geometries. The central claim is that a compact VAE latent representation enables robust assessment via evaluation of reconstruction accuracy for vortex generation across spatial and temporal scales, with emphasis on rear-end flow behavior.

Significance. If the VAE latent space demonstrably preserves geometry-specific vortex dynamics for both reconstruction and downstream prediction, the work would advance parametric model reduction for industrial aerodynamics, enabling efficient design-space exploration with reduced CFD cost. The distributed parallel training framework for high-resolution data is a noted practical strength.

major comments (2)
  1. Abstract: The evaluation is described solely in terms of reconstruction accuracy of vortex generation. However, the title specifies 'Predicting Turbulent Flow' and the prior work incorporated a time-evolution method; it is not shown whether the VAE latent variables support accurate time-stepping predictions (e.g., wake statistics or drag evolution) rather than only snapshot reconstruction. This distinction is load-bearing for the robustness claim across geometries.
  2. Abstract: No quantitative error metrics, comparisons to the non-VAE baseline, or details on latent dimension choice versus geometry variation are provided. Without these, it cannot be assessed whether the compact latent representation retains essential high-Re vortex features or merely averages over geometry-induced differences in separation and shedding.
minor comments (1)
  1. The abstract would be strengthened by specifying the range of Reynolds numbers, number of vehicle geometries, and any quantitative thresholds used to judge 'robustness'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each point below and will make revisions to improve clarity and completeness.

read point-by-point responses

  1. Referee: Abstract: The evaluation is described solely in terms of reconstruction accuracy of vortex generation. However, the title specifies 'Predicting Turbulent Flow' and the prior work incorporated a time-evolution method; it is not shown whether the VAE latent variables support accurate time-stepping predictions (e.g., wake statistics or drag evolution) rather than only snapshot reconstruction. This distinction is load-bearing for the robustness claim across geometries.

    Authors: We agree that the abstract focuses on reconstruction accuracy while the title refers to prediction, and that the prior work included time-evolution. The present study specifically evaluates the VAE extension for parametric robustness across geometries via reconstruction of vortex dynamics in the latent space; explicit time-stepping predictions with the new latent variables are not demonstrated here. We will revise the abstract to clarify this scope, note that time-evolution builds on the prior framework, and indicate that downstream prediction tasks remain for future work. revision: yes

  2. Referee: Abstract: No quantitative error metrics, comparisons to the non-VAE baseline, or details on latent dimension choice versus geometry variation are provided. Without these, it cannot be assessed whether the compact latent representation retains essential high-Re vortex features or merely averages over geometry-induced differences in separation and shedding.

    Authors: The referee correctly observes that the abstract lacks quantitative metrics, baseline comparisons, and latent-dimension details. The full manuscript reports reconstruction results across geometries, but the abstract does not summarize them. We will revise the abstract to include key quantitative error metrics, comparisons to the non-VAE baseline, and information on the selected latent dimension relative to geometry variation. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical extension of prior NN reduction via VAE reconstruction metrics

full rationale

The manuscript extends an earlier neural-network model-reduction framework by adding a variational autoencoder and reports reconstruction accuracy for vortex structures across geometries and Reynolds numbers. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text; the claims rest on direct numerical evaluation of reconstruction error rather than any derivation that reduces to its own inputs by construction. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract alone.

pith-pipeline@v0.9.1-grok · 5744 in / 951 out tokens · 31979 ms · 2026-06-25T22:06:41.126877+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 1 canonical work pages

  1. [1]

    C. Kato, Y. Yamade, K. Nagano, K. Kumahata, K. Minami, T. Nishikawa, ‘Toward Realization of Numerical Towing-Tank Tests by Wall-Resolved Large Eddy Simulation based on 32 Billion Grid Finite-Element Computation’, in: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, 2020, pp. 23–35

    2020

  2. [2]

    Available: https://www.fujitsu.com/global/documents/solutions/business-technology/tc/catalog/20180821hotchips30.pdf

    [Online]. Available: https://www.fujitsu.com/global/documents/solutions/business-technology/tc/catalog/20180821hotchips30.pdf

    arXiv

  3. [3]

    Available: https://www.r-ccs.riken.jp/en/fugaku/about/ (Accessed 12 September 2023)

    [Online]. Available: https://www.r-ccs.riken.jp/en/fugaku/about/ (Accessed 12 September 2023)

    2023

  4. [4]

    Jansson, M

    N. Jansson, M. Karp, A. Perez, T. Mukha, Y. Ju, J. Liu, S. Páll, E. Laure, T. Weinkauf, J. Schumacher, P. Schlatter, S. Markidis, ‘Exploring the Ultimate Regime of Turbulent Rayleigh–Bénard Convection Through Unprecedented Spectral-Element Simulations’, in: Proceedings of the International Conference for High Performance Computing, Networking, Storage and...

    2023

  5. [5]

    Available: https://www.lumi-supercomputer.eu/lumis-full-system-architecture-revealed/ (Accessed 14 February 2025)

    [Online]. Available: https://www.lumi-supercomputer.eu/lumis-full-system-architecture-revealed/ (Accessed 14 February 2025)

    2025

  6. [6]

    Vasudev, R

    K.L. Vasudev, R. Sharma, S.K. Bhattacharyya, ‘A multi-objective optimization design framework integrated with CFD for the design of AUVs’, Meth. Oceanogr. 10 (2014) 138–165

    2014

  7. [7]

    Kasagi, Y

    N. Kasagi, Y. Suzuki, K. Fukagata, Microelectromechanical systems–based feedback control of turbulence for skin friction reduction, Annu. Rev. Fluid Mech. 41 (2009) 231–251

    2009

  8. [8]

    Murata, K

    T. Murata, K. Fukami, K. Fukagata, ‘Nonlinear mode decomposition with convolutional neural networks for fluid dynamics’, J. Fluid Mech. 882 (2020) A13. https://doi.org/10.1017/jfm.2019.822

    work page doi:10.1017/jfm.2019.822 2020

  9. [9]

    K. Ando, K. Onishi, R. Bale, A. Kuroda, M. Tsubokura, ‘Nonlinear reduced-order modeling for three-dimensional turbulent flow by large-scale machine learning’, Comput. Fluids 266 (2023) 106047

    2023

  10. [10]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, ‘Long Short-Term Memory’, Neural Comput. 9 (1997) 1735–1780

    1997

  11. [11]

    B. Koo, H. Kim, T. Jo, S. Kim, J.Y. Yoon, ‘Proper orthogonal decomposition–Galerkin projection method for quasi-two-dimensional laminar hydraulic transient flow’, J. Hydraul. Res. 59 (2021) 224–234

    2021

  12. [12]

    Benner, S

    P. Benner, S. Gugercin, K. Willcox, ‘A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems’, SIAM Rev. 57 (2015) 483–531

    2015

  13. [13]

    J. Tran, K. Fukami, K. Inada, D. Umehara, Y. Ono, K. Ogawa, K. Taira, Aerodynamics-guided machine learning for design optimization of electric vehicles, Commun Eng 3 (2024)

    2024

  14. [14]

    Hasegawa, K

    K. Hasegawa, K. Fukami, T. Murata, K. Fukagata, ‘Machine-learning-based reduced-order modeling for unsteady flows around bluff bodies of various shapes’, Theor. Comput. Fluid Dyn. 34 (2020) 367–383

    2020

  15. [15]

    Hasegawa, K

    K. Hasegawa, K. Fukami, T. Murata, K. Fukagata, ‘CNN-LSTM based reduced order modeling of two-dimensional unsteady flows around a circular cylinder at different Reynolds numbers’, Fluid Dyn. Res. 52 (2020) 065501

    2020

  16. [16]

    Higashida, K

    A. Higashida, K. Ando, M. Rüttgers, A. Lintermann, M. Tsubokura, ‘Robustness evaluation of large-scale machine learning-based reduced order models for reproducing flow fields’, Future Gener. Comput. Syst. 159 (2024) 243–254

    2024

  17. [17]

    K. Ando, K. Onishi, R. Bale, A. Kuroda, M. Tsubokura, ‘Evaluation of applicability of nonlinear reduced-order model using neural network’, JSFM Annual Meeting 2022, 27-29 September, 2022, Kyoto, Japan (2022). (Japanese only)

    2022

  18. [18]

    Jansson, R

    N. Jansson, R. Bale, K. Onishi, M. Tsubokura, ‘CUBE: A scalable framework for large-scale industrial simulations’, Int. J. High Perform. Comput. Appl. 33 (2019) 678–698

    2019

  19. [19]

    Nakahashi, ‘Building-Cube method for flow problems with broadband characteristic length’, in: Computational Fluid Dynamics 2002, Springer, Berlin, Heidelberg, 2003, pp

    K. Nakahashi, ‘Building-Cube method for flow problems with broadband characteristic length’, in: Computational Fluid Dynamics 2002, Springer, Berlin, Heidelberg, 2003, pp. 77–81

    2002

  20. [20]

    Onishi, M

    K. Onishi, M. Tsubokura, ‘Topology-free immersed boundary method for incompressible turbulence flows: An aerodynamic simulation for “dirty” CAD geometry’, Comput. Methods Appl. Mech. Engrg. 378 (2021) 113734

    2021

  21. [21]

    Peskin, ‘The immersed boundary method’, Acta Numer

    C.S. Peskin, ‘The immersed boundary method’, Acta Numer. 11 (2002) 479–517

    2002

  22. [22]

    Bhalla, R

    A.P.S. Bhalla, R. Bale, B.E. Griffith, N.A. Patankar, ‘A unified mathematical framework and an adaptive numerical method for fluid-structure interaction with rigid, deforming, and elastic bodies’, J. Comput. Phys. 250 (2013) 446–476

    2013

  23. [23]

    Rüttgers, J

    M. Rüttgers, J. Park, D. You, ‘Large-eddy simulation of turbulent flow over the DrivAer fastback vehicle model’, J. Wind Eng. Ind. Aerodyn. 186 (2019) 123–138

    2019

  24. [24]

    Available: https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet (Accessed 27 February 2025)

    [Online]. Available: https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet (Accessed 27 February 2025)

    2025