Deep learning based on PINN for solving 2 D0F vortex induced vibration of cylinder with high Reynolds number

Cheng Chen, Pengfei Xu, Yongzheng Li, Guangtao ZhangarXiv 2021 · DOI · arXiv

Abstract

Vortex-induced vibration (VIV) exists widely in natural and industrial fields. The main approaches for solving VIV problems are numerical simulations and experimental methods. However, experiment methods are difficult to obtain the whole flow field information and also high-cost while numerical simulation is extraordinary time-consuming and limited in low Reynolds number and simple geometric configuration. In addition, numerical simulations are difficult to handle the moving mesh technique. In this paper, physics informed neural network (PINN) is proposed to solve the VIV and wake-induced vibration (WIV) of cylinder with high Reynolds number. Compared to tradition data-driven neural network, the Reynolds Average Navier-Stokes (RANS) equation, by implanting an additional turbulent eddy viscosity, coupled with structure's dynamic motion equation are also embedded into the loss function. Training and validation data is obtained by computational fluid dynamic (CFD) technique. Three scenarios are proposed to validate the performance of PINN in solving VIV and WIV of cylinders. In the first place, the stiffness parameter and damping parameter are calculated via limited force data and displacement data; secondly, the flow field and lifting force/drag force are inferred by scattered velocity information; eventually, the displacement can be directly predicted only through lifting forces and drag forces based on LSTM. Results demonstrate that,compared with traditional neural network, PINN method is more effective in inferring and re-constructing the unknown parameters and flow field with high Reynolds number under VIV and WIV circumstances.

Introduction

Vortex-induced vibration (VIV) occurs when fluid flow past a bluff body, such as a subsea pipeline or bridge pylon, generates alternating vortices. These vortices create periodic forces that can cause the structure to vibrate. When the frequency of these vibrations aligns with the structure's natural frequency, resonance occurs, potentially leading to catastrophic fatigue failure.

Traditional methods for analyzing VIV face significant hurdles. Experimental setups are costly and often fail to capture comprehensive flow field data. Conversely, numerical simulations using Computational Fluid Dynamics (CFD) are computationally expensive, particularly for high Reynolds number turbulent flows involving moving meshes. While standard deep learning offers a data-driven alternative, it typically requires vast amounts of data and lacks physical interpretability, acting as a "black box." This paper introduces a Physics-Informed Neural Network (PINN) framework to bridge these gaps, enabling efficient and physically consistent modeling of 2-Degree-of-Freedom (2 DOF) vibrations.

Methodology

The proposed framework leverages the PINN architecture to integrate physical laws directly into the learning process. Unlike standard networks that minimize only data-driven error, PINNs incorporate a "physics loss" term derived from governing equations.

Governing Equations and Turbulence Modeling

The authors embedded the Reynolds-Averaged Navier-Stokes (RANS) equations for fluid flow and the mass-spring-damper equations for structural motion into the network's loss function. To address the complexities of high Reynolds numbers, a learnable turbulent eddy viscosity ($\nu_t$) parameter was introduced into the RANS equations. This allows the network to dynamically adapt to turbulent structures.

Neural Network Architectures

The framework employs two primary architectures tailored for different tasks:

  1. Fully Connected Neural Networks (FCNN): Used for spatial flow reconstruction, mapping spatial coordinates and time to fluid properties.
  2. Long Short-Term Memory (LSTM): Integrated with 4th-order Runge-Kutta schemes for time-series trajectory prediction.
Architecture of the Fully Connected Neural Network (FCNN) used for spatial flow reconstruction.
Structure of the Long Short-Term Memory (LSTM) cell used for time-series trajectory prediction.

Automatic Differentiation (AD) via TensorFlow is utilized to compute the complex gradients required by the Navier-Stokes equations, bypassing the need for manual or numerical derivatives.

Data Generation

Training and validation data were synthesized using the pimpleDyMFoam solver within the OpenFOAM CFD framework. Two primary scenarios were investigated:

  • Single Cylinder (VIV): Conducted at $Re = 2889$.
  • Tandem Cylinders (Wake-Induced Vibration - WIV): Conducted at $Re = 9000$.

Latin Hypercube Sampling (LHS) was employed to select approximately 40,000 to 45,000 points for flow reconstruction. Notably, only 120 samples were required for structural parameter inference (stiffness and damping), highlighting the data efficiency of the PINN approach.

Discussion and Results

The PINN framework demonstrated high precision across several validation tasks:

  • Inverse Modeling: The model accurately identified structural parameters, predicting stiffness ($k$) with 0.62% error and damping ($c$) with a mere 0.02% error using minimal data.
  • Flow and Pressure Inference: A significant finding was the network's ability to reconstruct the entire pressure field accurately without any pressure data during training. By satisfying the momentum balance of the Navier-Stokes equations, the network "discovered" the underlying pressure distribution.
  • Force and Trajectory Prediction: Lift and drag forces were inferred from scattered velocity data with Mean Square Errors (MSE) between $10^{-4}$ and $10^{-5}$. The model also successfully captured the complex "butterfly" trajectories characteristic of vibrating cylinders.
  • Handling WIV: The framework proved robust in the more chaotic wake-induced vibration scenario, where a downstream cylinder interacts with the wake of an upstream one.

Conclusion and Future Work

This research demonstrates that PINNs are more effective than traditional neural networks for reconstructing unknown parameters and flow fields in high Reynolds number VIV and WIV scenarios. However, several avenues for future work remain:

  • Dimensionality: Extending the model from 2D to 3D to capture spanwise correlations inherent in real-world VIV.
  • Scaling: Increasing the Reynolds number to industrial scales ($Re > 10^5$).
  • Experimental Integration: Applying the framework to noisy, sparse data from experimental sensors or Particle Image Velocimetry (PIV) to further validate its real-world utility.

Summary updated

Abstract

Vortex-induced vibration (VIV) exists widely in natural and industrial fields. The main approaches for solving VIV problems are numerical simulations and experimental methods. However, experiment methods are difficult to obtain the whole flow field information and also high-cost while numerical simulation is extraordinary time-consuming and limited in low Reynolds number and simple geometric configuration. In addition, numerical simulations are difficult to handle the moving mesh technique. In this paper, physics informed neural network (PINN) is proposed to solve the VIV and wake-induced vibration (WIV) of cylinder with high Reynolds number. Compared to tradition data-driven neural network, the Reynolds Average Navier-Stokes (RANS) equation, by implanting an additional turbulent eddy viscosity, coupled with structure's dynamic motion equation are also embedded into the loss function. Training and validation data is obtained by computational fluid dynamic (CFD) technique. Three scenarios are proposed to validate the performance of PINN in solving VIV and WIV of cylinders. In the first place, the stiffness parameter and damping parameter are calculated via limited force data and displacement data; secondly, the flow field and lifting force/drag force are inferred by scattered velocity information; eventually, the displacement can be directly predicted only through lifting forces and drag forces based on LSTM. Results demonstrate that,compared with traditional neural network, PINN method is more effective in inferring and re-constructing the unknown parameters and flow field with high Reynolds number under VIV and WIV circumstances.
BibTeX
@misc{chen2021deep,
  author = {Cheng Chen and Pengfei Xu and Yongzheng Li and Guangtao Zhang},
  title = {{Deep learning based on PINN for solving 2 D0F vortex induced vibration of cylinder with high Reynolds number}},
  year = {2021},
  eprint = {2106.01545},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.01545},
}
Figure