Prediction of turbulent energy based on low-rank resolvent modes and machine learning

Yitong Fan, Bo Chen, Weipeng LiJournal of Physics Conference Series 2024 · DOI

TL;DR

This paper introduces a novel framework that combines resolvent analysis with machine learning to predict turbulent energy in channel flows. By leveraging low-rank resolvent modes as basis functions and training neural networks to learn their coefficients, the method accurately reconstructs turbulent energy distributions, offering a computationally efficient alternative to expensive high-fidelity simulations.

What Problem Does This Paper Address?

Understanding and predicting turbulent flows is crucial for many engineering applications, from designing aircraft to controlling fluid dynamics. However, high-fidelity numerical simulations like Direct Numerical Simulation (DNS) are computationally prohibitive. Traditional methods like Reynolds-Averaged Navier-Stokes (RANS) offer computational efficiency but lack accuracy in predicting turbulent fluctuations and structures. Reduced-order models (ROMs) like Proper Orthogonal Decomposition (POD) and Spectral POD (SPOD) are data-driven and can capture key flow features, but often require extensive post-processing. This paper aims to bridge this gap by developing a physics-informed, data-driven approach that efficiently predicts turbulent energy distributions using a low-rank representation derived from resolvent analysis.

Methods

The core of this work lies in a hybrid approach:

  • Resolvent Analysis: This method linearizes the Navier-Stokes equations around a steady base flow, creating an input-output system that relates forcing terms to flow responses. By applying Singular Value Decomposition (SVD) to the resolvent operator, the authors extract dominant "resolvent modes." These modes represent the most amplified linear responses to specific forcing frequencies and wavenumbers and are hypothesized to capture the low-rank behavior of turbulent flows.
  • Machine Learning (Convolutional Neural Networks - CNNs): The coefficients (a_p) that weight these resolvent modes to reconstruct the full flow field are learned using a CNN. The input to the CNN is the mean streamwise velocity profile, which is a readily available quantity from lower-fidelity simulations like RANS. The output of the CNN is the set of projection coefficients for the resolvent modes.
  • Loss Function: The CNN is trained by minimizing a Mean Squared Error (MSE) loss function. Two variations are explored: one that penalizes only the streamwise energy prediction error, and another that penalizes errors across all velocity and pressure components.

Datasets

The study utilizes data from Direct Numerical Simulations (DNS) for incompressible turbulent channel flows at various friction Reynolds numbers ($Re_\tau$): 180, 390, 550, 1000, 2000, 4200, and 5200. These DNS datasets provide the ground truth for the mean base flow (used as input to the CNN) and the turbulent energy distributions (used for training and validation). The friction Reynolds number of 1000 is specifically used for testing the trained model.

Key Insights and Ideas

  • Low-Rank Structure of Turbulence: The paper reinforces the idea that turbulent flows, despite their apparent complexity, exhibit a significant low-rank structure that can be effectively captured by a limited number of resolvent modes.
  • Physics-Informed Machine Learning: By using resolvent modes derived from the linearized Navier-Stokes equations as basis functions, the machine learning model is inherently grounded in the physics of fluid dynamics, leading to more robust and generalizable predictions.
  • Mean Flow as a Powerful Input: The finding that the mean streamwise velocity profile alone is sufficient input for the CNN to predict turbulent energy distributions highlights the strong correlation between the mean flow and the turbulent structures.
  • Trade-offs in Loss Function Design: The choice of loss function significantly impacts the prediction accuracy for different flow components. Penalizing only streamwise energy leads to excellent streamwise energy prediction but poorer accuracy for other components, suggesting that the optimal loss function depends on the specific prediction target.

Key Results

  • Accurate Streamwise Energy Prediction: The framework demonstrates highly accurate prediction of the streamwise turbulent energy distribution, with Mean Squared Error (MSE) values often below 10% for various Reynolds numbers, especially when the loss function prioritizes streamwise energy.
  • Reasonable Prediction of Other Components: While less accurate than streamwise energy, the prediction of wall-normal and spanwise velocity components and pressure perturbation shows reasonable agreement, particularly when the loss function considers all components.
  • Capturing Inner-Layer Structures: The model effectively captures the inner-layer peaks in premultiplied spanwise spectra, which are representative of near-wall streaks and streamwise vortices.
  • Computational Efficiency: The proposed method offers a significant computational advantage over DNS, enabling predictions with readily available mean flow data.

Limitations and Future Work

  • Outer-Layer Discrepancies: The prediction error tends to increase for larger scales in the outer layer, especially at high Reynolds numbers. This is attributed to the limited availability of DNS data for these scales and the inherent limitations of the rank-1 resolvent approximation.
  • Database Dependency: The accuracy of the predictions is influenced by the size and quality of the training databases.
  • Future Work: The authors suggest exploring the effect of increasing the degrees of freedom of the projection coefficients and the number of resolvent-based modes used. Additionally, incorporating more databases for training and testing is expected to further improve predictive accuracy. Investigating the prediction of instantaneous three-dimensional flow fields is also mentioned as a potential future direction.

Summary updated

Abstract

Abstract A modelling framework based on the resolvent analysis and machine learning is proposed to predict the turbulent energy in incompressible channel flows. In the framework, the optimal resolvent response modes are selected as the basis functions modelling the low-rank behaviour of high-dimensional nonlinear turbulent flow-fields, and the corresponding weight functions are determined by data-driven neural networks. Turbulent-energy distribution in space and scales, at the friction Reynolds number 1000, is predicted and compared to the data of direct numerical simulation. Close agreement is observed, suggesting the feasibility and reliability of the proposed framework for turbulence prediction.
BibTeX
@article{fan2024prediction,
  author = {Yitong Fan and Bo Chen and Weipeng Li},
  title = {{Prediction of turbulent energy based on low-rank resolvent modes and machine learning}},
  journal = {Journal of Physics Conference Series},
  year = {2024},
  doi = {10.1088/1742-6596/2753/1/012021},
}
Figure