Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

Navier-Stokes equations are dissipative, meaning the long-time dynamics of a flow can be described in a low-dimensional coordinate system.
Data-driven manifold dynamics modeling method can be used to describe the dynamics of turbulent Couette flow.
Autoencoder used to find low-dimensional manifold coordinate system and neural network used to define ordinary differential equations.
Models with fewer than 20 degrees of freedom can capture key characteristics of the flow.
Models outperform POD-Galerkin models with 2000 degrees of freedom.
Unstable periodic orbits computed from the models, some of which correspond to previously unknown solutions.

Major challenge in dealing with chaotic fluid flows is high-dimensional nature of the state
Dissipative nature of turbulent flows leads to expectation that long-time dynamics collapse onto an invariant manifold of much lower dimension
Modeling dynamics in a manifold coordinate system can speed up computations and be useful for downstream tasks
Proper orthogonal decomposition (POD) is a linear dimension reduction technique
POD has been used for many flow phenomena
Nonlinear methods can be used for more accurate reconstruction
Autoencoders can be used with equations of motion in manifold Galerkin approach
Machine learning techniques can be used to approximate dynamics
Neural ordinary differential equations (ODE) can be used to model chaotic dynamics
Aim is to use data-driven model to reconstruct turbulent plane Couette flow
Self-sustaining process (SSP) of wall bounded turbulence outlined by Hamilton et al. (1995)
Low-dimensional models have been developed to recreate the dynamics of the SSP
Data-driven models have been used to attempt to reconstruct the MFE model dynamics
Key approach to understanding complex nonlinear dynamical phenomena is through study of underlying state space structure of fixed points and periodic orbits
Data-driven Manifold Dynamics (DManD) used to build a ROM for turbulent MFU PCF

DManD modeling approach allows for perfect reconstruction
Training data is full-state data in an ambient space
Mapping to low-dimensional state representation with ℎ
Mapping back to full state with hybrid autoencoder
Autoencoder trained by minimizing mean-squared error and promoting accurate representation of POD coefficients
Evolution of trajectories on manifold performed in manifold coordinates
Approximate with neural ODE and linear damping term
Determine parameters by minimizing difference between predicted and true state
Preprocessing step to split state into pattern and phase
Evolution of pattern and phase only depend on pattern
Additional ODE for phase
Determine parameters of phase ODE by minimizing difference between predicted and true phase

Applied DManD to DNS of turbulent PCF in a MFU domain
Considered the well-studied Hamilton, Kim, and Waleffe (HKW) domain
Compared DManD results to analysis of self-sustaining process in this domain
Compared DManD results to other Galerkin-based ROMs
Compared DManD results to known unstable periodic solutions in this domain
Solved Navier-Stokes equations for fluid confined between two plates moving in opposite directions
Selected HKW cell because it is well studied and isolates the SSP
Generated turbulent PCF trajectories using psuedo-spectral Channelflow code
Split data into 80% for training and 20% for testing
Preprocessed data by computing mean u and subtracting it from all data

Linear dimension reduction with POD reduces the dimension from O (10 5 ) to 502
POD modes are treated as Fourier modes in both the and directions
256 POD modes contain 99.8% of the energy
Autoencoders are used to reduce the dimension down to ℎ
Autoencoders perform an order magnitude better than POD
Stabilized neural ODEs are trained to predict the evolution of the centered data
NN is trained with an Adam optimizer and learning rate scheduler

DManD models are evaluated for reconstructing short-time trajectories and long-time statistics
DManD captures key characteristics of the SSP (low-speed streaks become wavy, rolls lift fluid from the walls, regenerating streaks)
Quantitative agreement between DNS and DManD
Modal root-mean squared velocity used to identify different parts of the SSP
Tracking error of 10 trajectories shown in Fig. 5
Normalized ensemble-averaged tracking error for model dimensions between 3 and 18
Instantaneous kinetic energy of the flow is considered
ODE trained for phase dynamics
Mean squared phase displacement in good agreement between DNS and model

DManD model uses stabilized neural ODEs with an explicit damping term to capture long-time dynamics
Without damping, almost all models become unstable for ℎ = 5-16, and around half become unstable in other cases
DManD model with 18 degrees of freedom matches mean-squared amplitudes to high accuracy
For ℎ = 2 and 3, nearly all models match the data, with small deviations for ℎ ∼ 3-6
For ℎ = 2 and 2, deviation becomes more obvious, and lines do not converge until around ℎ > 10
With ℎ = 6, model overestimates number of low dissipation states, while ℎ = 18 matches density well
EMD steeply drops at ℎ = 5 and after ℎ > 10 joint PDFs are in excellent agreement with DNS

Data-driven manifold dynamics method (DManD) used to model MFU PCF with fewer degrees of freedom than required for DNS
Hybrid autoencoder approach used to find low-dimensional parameterization of manifold
Stabilized neural ODEs used to accurately evolve low-dimensional state forward in time
DManD model captures self-sustaining process and accurately tracks trajectories and temporal autocorrelation
POD-Galerkin model requires 2000 degrees of freedom to match mean-squared POD coefficient amplitudes
DManD model used for PO searches, successfully landed on nearby POs and found 9 previously undiscovered RPOs
True dimension of turbulent flow dynamics can be orders of magnitude smaller than degrees of freedom required for fully-resolved simulation
DManD model enables model-based nonlinear control algorithms to be used
Research supported by AFOSR FA9550-18-1-0174 and ONR N00014-18-1-2865