Moehlis–Faisst–Eckhardt Turbulence Model

• The MFE model is a nine-dimensional low-order turbulence model defined by a Galerkin projection of the Navier–Stokes equations onto divergence-free modes, capturing key flow structures.
• It formulates nine coupled nonlinear ODEs that incorporate linear decay, quadratic mode interactions, and external forcing to model the self-sustaining process of wall turbulence.
• The model serves as a benchmark for chaos, bifurcation analysis, and data-driven approaches, including machine learning, to predict extreme turbulent events.

The Moehlis–Faisst–Eckhardt (MFE) model is a paradigmatic nine‐dimensional low‐order model for turbulent shear flows, constructed via Galerkin projection of the incompressible Navier–Stokes equations onto a specifically chosen set of nine divergence‐free spatial modes. Designed to minimally capture the self‐sustaining process of wall‐bounded turbulence—including the interplay of rolls, streaks, and three‐dimensional instabilities—the MFE model provides a benchmark for understanding nonlinear transitions, coherent structures, rare extreme events, and for developing data‐driven prediction and reduction frameworks in turbulence research. Its mathematical structure, bifurcation scenarios, and tractable dimensionality have made it a cornerstone in studies of chaos, model reduction, and machine‐learning-based forecasting of rare turbulent events.

1. Mathematical Formulation

The MFE model consists of nine coupled nonlinear ODEs for the modal amplitudes $a_j(t)$ ($j = 1,\dots, 9$), representing the velocity field in a three-dimensional domain with free-slip or no-slip walls and periodicity in the streamwise and spanwise directions. The ODEs are obtained by Galerkin projection of the Navier–Stokes equations with a body force that sustains a laminar base profile. In canonical quadratic form, the dynamics are: $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$

• $L_{ij}$: linear coefficients representing viscous decay and linear couplings;
• $N_{ijk}$: all triadic nonlinear interaction coefficients derived from the convective term;
• $F_i$: external forcing, nonzero only for the mean-flow mode ($i=1$).

Parameter values are typically set for the turbulent regime; e.g., $\text{Re} = 400$, domain sizes $(L_x, L_y, L_z) = (4\pi, 2, 2\pi)$, with fundamental wavenumbers $\alpha = 2\pi/L_x = 0.5$, $j = 1,\dots, 9$0, $j = 1,\dots, 9$1 (Golyska et al., 2023, 2002.01222).

Boundary conditions are encoded via the choice of basis: periodic in $j = 1,\dots, 9$2 and $j = 1,\dots, 9$3, free-slip or no-slip in $j = 1,\dots, 9$4 depending on the variant (Constante-Amores et al., 2024).

2. Physical Interpretation and Modal Basis

Each mode amplitude $j = 1,\dots, 9$5 corresponds to a distinct spatial structure essential for the sustenance and instability of wall-bounded turbulence:

• $j = 1,\dots, 9$6: mean-flow correction (deviation from laminar base flow);
• $j = 1,\dots, 9$7: streamwise vortices (rolls) responsible for streak generation;
• $j = 1,\dots, 9$8: streak-associated wave modes (sinuous/varicose), become unstable on developed streaks;
• $j = 1,\dots, 9$9: secondary oblique/3D modes that feedback to regenerate rolls and mean distortion, closing the self-sustaining cycle (Golyska et al., 2023, Srinivasan et al., 2019, 2002.01222).

All quadratic mode couplings required by the underlying symmetries and momentum transfer mechanisms are included, and the resulting nine-dimensional state space is sufficient to support key turbulent phenomena.

3. Dynamical Phenomena: Attractors, Regimes, and Invariant Solutions

The MFE model possesses a nine-dimensional phase space structured by multiple invariant sets:

• The trivial fixed point, $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$0, corresponds to the laminar solution. Its stable manifold separates transient decay to laminar flow from sustained turbulence.
• For $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$1, the system supports a robust chaotic attractor, with sustained aperiodic fluctuations in energy and dissipation representative of wall-bounded turbulence.
• Embedded within this attractor are unstable periodic orbits, edge states, and quasi-laminarization bursts ("hibernation events") (Golyska et al., 2023, Constante-Amores et al., 2024).
• Global bifurcations, such as saddle-node and Hopf bifurcations, orchestrate the emergence of nontrivial steady solutions and complex transient behavior. For $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$2, there are precisely eight stationary solutions, falling into two symmetry-related families, found via Gröbner basis elimination of the polynomial fixed-point equations (Pausch et al., 2014).

These solutions and their bifurcations have been studied both analytically (polynomial solution structure) and numerically, elucidating pathways to turbulence and rare event statistics.

4. Diagnostics, Statistical Behavior, and Extreme Events

Standard diagnostics for the MFE model include:

• Turbulent kinetic energy $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$3
• Energy-dissipation rate $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$4, where $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$5 is a mode-projected dissipation matrix.

The model captures three principal regimes:

• Laminar decay ($$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$6, $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$7)
• Chaotic/turbulent state (statistically stationary with nontrivial fluctuations)
• Bursts/extreme events: intermittent, high-amplitude excursions in $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$8 and $$\frac{da_i}{dt} = \sum_{j=1}^9 L_{ij} a_j + \sum_{j=1}^9 \sum_{k=1}^9 N_{ijk} a_j a_k + F_i, \quad i = 1,\dots,9$$9; these are rare but crucial for capturing the fat tails of turbulent statistics and are associated with excursions toward the basin boundary and return to turbulence (Golyska et al., 2023, Fox et al., 2023).

Computation of Lyapunov exponents ($L_{ij}$0 for $L_{ij}$1), joint PDFs, return statistics, and survival functions for quasi-laminarization trajectories allow in-depth quantification of predictability, rare-event frequencies, and state-space geometry (Fox et al., 2023).

5. Analytical and Numerical Methods

The structure of the MFE model enables both analytical and numerical solution strategies:

• Fixed points: Systematic construction and classification via Gröbner basis reduction enables global enumeration of all stationary states, revealing eight real fixed points at $L_{ij}$2 (Pausch et al., 2014).
• Integration: Standard explicit Runge–Kutta schemes (RK4) are used, with time-steps $L_{ij}$3–$L_{ij}$4, large ensemble runs ($L_{ij}$5–$L_{ij}$6), and finely-resolved state sampling (Golyska et al., 2023, 2002.01222).
• Symmetry reduction: Exploitation of reflection and half-cell shift symmetries in $L_{ij}$7 and $L_{ij}$8 further reduces effective dimensionality of the accessible phase space (Pausch et al., 2014).

Such tractability permits exhaustive statistical analyses and reliable generation of surrogate datasets for downstream machine-learning tasks.

6. Data-Driven Modeling, Machine Learning, and Koopman Approaches

The MFE model has played a pivotal role as a testbed for developing and benchmarking data-driven prediction, model reduction, and extreme-event forecasting methodologies:

• Neural networks: Both multilayer perceptrons (MLP) and long short-term memory (LSTM) recurrent networks trained on MFE-generated time series yield mean and fluctuation errors $L_{ij}$9, with LSTM outperforming MLP due to the sequential structure of turbulence (Srinivasan et al., 2019, 2002.01222). Loss functions incorporating statistical averages accelerate training and improve performance.
• Clustering and charts/atlases: The CANDyMan framework builds a global dynamical atlas by clustering the phase space and deploying specialized neural maps ("charts"), dramatically reducing error in forecasting extreme events and improving tail statistics relative to single-network and standard extended dynamic mode decomposition (EDMD) approaches (Fox et al., 2023).
• Projected Koopman Dynamics: Data-driven approximation of the Koopman operator, using a neural-network-learned dictionary ($N_{ijk}$0 observables), achieves accurate short-term tracking ($N_{ijk}$1 Lyapunov times) and long-term statistical fidelity, outperforming both CANDyMan and neural-ODE models in both short-term and statistical metrics (Constante-Amores et al., 2024).
• Clustering-based precursor identification: State-space tessellation and clustering (modularity-based) provide probabilistic pathways for predicting extreme bursts by identifying and statistically characterizing precursor states (Golyska et al., 2023).

These approaches have established the MFE model as an indispensable benchmark for rigorous, quantitative, and comparative studies of data-driven turbulent flow prediction and rare-event analytics.

7. Impact and Extensions

The MFE model has established itself as a minimal yet dynamically rich laboratory for the study of wall-bounded shear flow phenomena, transition to chaos, and rare excursions fundamental to turbulence. Its compactness allows for:

• Full enumeration of invariant solutions and attractors, including fixed points, periodic orbits, and chaotic saddles.
• Systematic exploration of bifurcation structure and critical Reynolds numbers.
• Controlled benchmarking and validation of data-driven surrogates, with clear pathways for extension to higher-order truncations or alternative geometric settings.
• Theoretical insights into the role of coherent structures, symmetry breaking, and finite-dimensional representations of turbulence.

Extensions explored include modification of the mode set, refinement of forcing profiles, use in higher-fidelity datasets for extreme-event detection and prediction, and generalization to non-polynomial flows, with the model serving as a canonical reference against which newly developed methodologies are validated (Pausch et al., 2014, Constante-Amores et al., 2024, Golyska et al., 2023, Fox et al., 2023).