High-Efficient Machine Learning Projection Method for Incompressible Navier-Stokes Equations

Ruilin Chen, Xiaowei Jin, Nikolaus A. Adams, Hui LiSSRN 2025 · DOI

TL;DR

This paper introduces a novel machine learning (ML) projection method to accelerate the solution of incompressible Navier-Stokes equations, a fundamental challenge in fluid dynamics. By generating synthetic, high-fidelity data and embedding ML models into the traditional projection framework, the authors achieve significant speedups (up to 7x) without sacrificing accuracy, overcoming the bottleneck of expensive real-world data acquisition.

What Problem Does This Paper Address?

Solving the incompressible Navier-Stokes equations is crucial for simulating fluid flow, but traditional numerical methods are computationally expensive, especially for high-fidelity simulations. A key bottleneck is the pressure Poisson equation (PPE), which requires significant computational resources. While machine learning (ML) offers potential for acceleration, its application in this domain is hampered by the need for massive amounts of high-fidelity training data (e.g., from Direct Numerical Simulations or expensive experiments), which are often inaccessible or prohibitively costly. This paper aims to bridge this gap by developing an ML-based projection method that can be trained on synthetically generated data, thereby reducing the reliance on expensive real-world datasets.

Methods

The core of the proposed method is an "ML projection method" that integrates ML models into the existing projection framework for solving Navier-Stokes equations. The key innovation lies in how the ML models are trained and utilized:

  • Data Forward Generation: Instead of relying on expensive real-world data, the authors generate synthetic high-fidelity data. This is done by:
  • Gaussian Random Fields (GRF): Generating flow fields that follow statistical properties of Gaussian distributions.
  • Spectra Constraint: Generating flow fields that adhere to specific spectral properties (e.g., turbulent kinetic energy spectra), which are more physically representative of certain flow phenomena.
  • PDE Balance: Once a flow field is generated, the source terms, boundary conditions, and initial conditions are forward-computed to ensure they satisfy the governing partial differential equations (PDEs). This creates consistent data pairs (inputs and outputs) for training.
  • ML Embedded Solvers: Two specific ML models are embedded into the projection method to solve the computationally expensive PPE:
  • Poisson Neural Network (Poisson-NN): This network is designed to directly approximate the solution of the PPE. It can either provide a good initial guess for a traditional multigrid (MG) solver or, if accurate enough, replace it entirely.
  • Wavelet Transform Convolutional Neural Network Multigrid (WTCNN-MG): This approach integrates a CNN with a multigrid framework, leveraging wavelet transforms. It aims to optimize the smoothing, differentiation, restriction, and prolongation operations within the MG solver, and also incorporates low-frequency error correction.

The ML models are integrated into the numerical framework in either a "patchwork" manner (Poisson-NN) or an "overall differentiable" manner (WTCNN-MG).

Datasets

The paper does not use pre-existing high-fidelity datasets like DNS data for training. Instead, it focuses on generating its own synthetic datasets:

  • Generated Datasets:
    • GRF-based: Pressure fields generated using Gaussian Random Fields with Matérn kernel parameters.
  • Spectra-constraint-based: Pressure fields generated using specific spectral constraints (e.g., von Kármán-Pao spectrum).
  • These datasets consist of pressure fields and their corresponding right-hand side terms (source terms) for the PPE.

The generated datasets are used to train the Poisson-NN and WTCNN-MG models. The performance of these ML models is then evaluated on standard fluid dynamics problems (Kolmogorov flows and decaying flows) where the ground truth is obtained from traditional numerical solvers.

Key Insights and Ideas

  • Data Generation is Key: The ability to generate realistic, high-fidelity training data synthetically is a critical enabler for ML in CFD. This bypasses the expensive data acquisition bottleneck.
  • Spectra Constraint > GRF: Generating data based on physical spectra (like turbulent kinetic energy spectra) leads to better generalization and accuracy, especially for high-frequency components, compared to purely statistical GRF generation. This highlights the importance of incorporating physical constraints into synthetic data.
  • ML as a Solver/Preconditioner: ML models can effectively act as either direct solvers for the PPE or as highly efficient preconditioners for traditional solvers, significantly reducing iteration counts.
  • WTCNN-MG's Superiority: The WTCNN-MG approach, by combining wavelet transforms and CNNs within a multigrid framework, shows superior performance in terms of convergence speed and accuracy compared to simpler ML models or traditional MG.
  • Double Precision Matters: Switching to double precision during training significantly improves the final loss, indicating that single precision can be a limiting factor for achieving high accuracy in complex ML models.

Key Results

  • Significant Speedups:
    • The Poisson-NN embedded method achieved an average speed-up of 5.83 times over the traditional MG method.
  • The WTCNN-MG method achieved an even greater average speed-up of 7.03 times.
  • These speedups were observed in both Kolmogorov flows and decaying flow simulations.
  • High Accuracy: The ML models, trained on synthetic data, achieved excellent generalization and accuracy, with relative errors consistently lower than the time-step error.
  • Spectra-Constraint Advantage: Spectra-constrained data led to better high-frequency approximation and overall accuracy compared to GRF-generated data. For example, in decaying flows, spectra-based Poisson-NN reduced relative errors by nearly two orders of magnitude compared to GRF-based.
  • WTCNN-MG Performance: WTCNN-MG consistently outperformed MG, NMG, and CNN-MG in terms of convergence rate and iteration counts, especially in decaying flow tasks and 3D isotropic flows. It achieved final residuals of 1x10⁻⁶ significantly faster than other methods.

Limitations and Future Work

  • High-Precision Requirements: While the ML models achieve significant speedups, they did not yet meet the extremely high precision requirement of 1x10⁻⁶ for all cases. They are thus used as excellent initial approximations for traditional solvers.
  • Generalization to Diverse Flows: The current study focuses on specific flow types (Kolmogorov, decaying, isotropic). Future work could explore anisotropic turbulent flows with separating, attaching, and transition phenomena.
  • Further Optimization of WTCNN-MG: While WTCNN-MG shows excellent results, further research into its architecture and training could yield even greater improvements.
  • Broader Application: The paper suggests that the ML projection method and data forward generation method have broad application potential beyond the tested scenarios.

Summary updated

BibTeX
@misc{chen2025high,
  author = {Ruilin Chen and Xiaowei Jin and Nikolaus A. Adams and Hui Li},
  title = {{High-Efficient Machine Learning Projection Method for Incompressible Navier-Stokes Equations}},
  year = {2025},
  doi = {10.2139/ssrn.5120877},
  howpublished = {SSRN},
}
Figure