Data-driven Symbolic Closure for Turbulence Modeling in the Lattice Boltzmann Framework

This paper presents a data-driven approach using Physical Symbolic Optimization to discover an interpretable, non-linear analytical closure for Lattice Boltzmann turbulence modeling that outperforms traditional Smagorinsky models in accuracy and generalizes robustly to wall-bounded flows without supplemental corrections.

The Gist

Imagine you are trying to predict how a cup of coffee swirls when you stir it, or how smoke curls from a candle. In the world of physics, this chaotic, twisting motion is called turbulence. It is one of the hardest puzzles in science because the fluid moves in tiny, unpredictable patterns that change constantly.

To simulate this on a computer, scientists use a method called Lattice Boltzmann. Think of this method as a giant grid of tiny tiles. Instead of tracking every single molecule of fluid, the computer looks at how "particles" hop from one tile to the next.

The Problem: The "Over-Blunt" Knife

The paper explains that when we try to simulate turbulence on a computer, we can't afford to make the grid so fine that it catches every tiny swirl (that would take too much computing power). So, we use a "shortcut" called a Sub-Grid Scale (SGS) model.

Think of the SGS model as a chef's knife used to chop vegetables.

• The Old Knife (Smagorinsky Model): For decades, scientists have used a standard model (Smagorinsky) that acts like a very blunt, heavy cleaver. It chops everything roughly the same way. Near walls (like the side of a pipe), this cleaver is too aggressive. It chops up the delicate, small swirls that should be there, making the simulation "over-dissipative" (it loses energy too fast) and missing important details like tiny corner vortices.
• The Goal: The researchers wanted a scalpel—a tool that knows exactly how hard to cut in different situations, preserving the delicate details without wasting energy.

The Solution: Teaching a Computer to Write the Recipe

Instead of trying to guess the perfect formula using old-school math theories, the authors used a "data-driven" approach. They used a technique called Physical Symbolic Optimization (Φ-SO).

Here is the analogy:
Imagine you have a massive library of high-definition videos showing exactly how fluids move (these are called DNS datasets). You want the computer to look at these videos and write down a simple mathematical "recipe" (an equation) that explains the movement.

Usually, computers use "Black Box" AI (like deep neural networks) to do this. They give you an answer, but you can't see how they got there. It's like a magic trick where you see the rabbit appear, but you don't know the trick.

This paper used a different approach:

1. The Search: The computer was given a toolbox of math symbols (plus, minus, multiply, square roots, etc.) and a set of rules based on physics (like "energy must scale in a certain way").
2. The Discovery: The computer tried millions of different combinations of these symbols, checking them against the high-definition videos. It kept the formulas that worked best and discarded the ones that were too complicated or didn't fit the physics.
3. The Result: It found a specific, readable equation (a "recipe") that acts like a smart scalpel.

What Makes This New Recipe Special?

The new formula the computer discovered is "smart" because it looks at two things at once:

1. Stretching (Strain): How much the fluid is being pulled apart.
2. Spinning (Rotation): How much the fluid is twisting.

The old "blunt cleaver" only looked at stretching. The new "scalpel" knows that if the fluid is spinning fast but not stretching much, it should behave differently. This allows it to:

• Preserve the delicate details: In a simulation of a box with a moving lid, the new model successfully found tiny, weak swirls in the corners (called Moffatt eddies) that the old model completely smoothed over and erased.
• Work without a manual override: The old models often needed a special "damping" rule added by hand to stop them from being too aggressive near walls. The new model figured this out on its own.

The "Zero-Shot" Magic Trick

The most impressive part of the paper is the generalization test.

• The computer was trained only on two specific types of flows: a swirling vortex in open space and a box with a moving lid.
• Then, the researchers asked it to simulate a completely different scenario: Turbulent flow in a pipe (channel flow), which it had never seen before.
• The Result: Without any extra training or "cheat codes" for pipes, the model performed better than the standard method. It correctly predicted how the fluid moves near the pipe walls, proving it learned a fundamental rule of turbulence rather than just memorizing the training videos.

Summary

In simple terms, the authors used a smart computer search to find a new, simpler, and more accurate mathematical rule for simulating turbulent fluids.

• Old way: Use a blunt tool that misses details and needs constant manual fixing.
• New way: Use a computer to discover a precise, self-correcting formula that understands both stretching and spinning, allowing it to see the "fine print" of turbulence that others miss.

This work suggests that in the future, we might not need to guess how fluids behave; we can let data-driven tools discover the laws of physics for us, creating smarter and more efficient simulations for engineering and science.

Technical Summary

Technical Summary: Data-driven Symbolic Closure for Turbulence Modeling in the Lattice Boltzmann Framework

1. Problem Statement

Turbulence modeling within the Lattice Boltzmann Method (LBM) framework has historically relied on traditional algebraic Sub-Grid Scale (SGS) models, most notably the Smagorinsky model. While effective in homogeneous isotropic turbulence, these models suffer from critical limitations in complex flow regimes:

• Over-dissipation: They tend to be overly dissipative near solid boundaries, damping out important flow structures.
• Lack of Spatial Selectivity: They struggle to distinguish between laminar shear and turbulent rotation, often introducing non-physical eddy viscosity in laminar regions.
• Stagnation in Derivation: The development of new purely algebraic SGS models has reached a bottleneck, with few breakthrough theories emerging from traditional derivation methods.
• Interpretability Gap: While data-driven Machine Learning (ML) approaches (e.g., CNNs, GANs) have shown high accuracy, their "black-box" nature limits physical interpretability and integration into standard CFD solvers without heavy dependencies.

The central challenge is to discover an explicit, interpretable, and physically consistent algebraic closure for the effective relaxation time ($\tau_{eff}$) in LBM-LES that overcomes the deficiencies of traditional models while avoiding the opacity of neural networks.

2. Methodology

The authors propose a data-driven framework utilizing Physical Symbolic Optimization (Φ-SO) to discover explicit analytical closures. The workflow integrates high-fidelity Direct Numerical Simulation (DNS) data with a symbolic regression algorithm enhanced by physical constraints.

A. Data Generation and Extraction

• Datasets: High-fidelity DNS data was generated using the open-source OpenLB code for two canonical flows:
  1. Taylor-Green Vortex (TGV): Homogeneous isotropic turbulence at $Re = 800, 1600, 3000$.
  2. Lid-Driven Cavity (LDC): Inhomogeneous wall-bounded turbulence at $Re = 3200$ and $10,000$.
• Filtering: To emulate Large Eddy Simulation (LES), a Gaussian spatial filter was applied to the DNS distribution functions.
• Target Variable: The effective relaxation time ($\tau_{eff}$) was extracted from the filtered LBM equation. The turbulent contribution, $\tau_{turb} = \tau_{eff} - \tau_0$, serves as the target variable for the regression.
• Feature Engineering: Input features were derived from the filtered velocity gradient tensor, including strain-rate ($S_{ij}$) and rotation-rate ($\Omega_{ij}$) tensors, along with their invariants ($I, II, III, IV$).

B. The Φ-SO Algorithm

The core discovery process employs the Φ-SO algorithm, which operates via a Reinforcement Learning loop:

1. Expression Generation: A Recurrent Neural Network (RNN) agent generates mathematical expression trees token by token.
2. Constant Optimization: Continuous parameters within candidate expressions are optimized using gradient-based methods to minimize Mean Squared Error (MSE).
3. Reward Function: A reward function balances accuracy (MSE) and simplicity (complexity penalty) to guide the search toward parsimonious models.

C. Physics-Informed Constraints

To ensure physical consistency, the authors introduced specific constraints into the symbolic search:

• Virtual Dimensional Analysis: Virtual physical dimensions were assigned to dimensionless LBM variables. The algorithm was constrained to enforce scaling laws (e.g., $\tau_{eff} \propto |S|$), pruning physically inconsistent expressions.
• Non-linear Tensor Invariants: The search space was expanded to include higher-order non-linear invariants (e.g., coupling rotation and strain) and fractional power operators (specifically pow(1/4)) to map high-order terms to the required target dimensionality.

3. Key Contributions

The paper presents three primary contributions:

1. Physics-Informed Discovery Workflow: The integration of virtual dimensional analysis and non-linear tensor invariants into symbolic regression. This strategy enforces physical scaling laws directly within the search process, enabling the discovery of a model that respects fundamental physics without manual derivation.
2. Interpretable & Explicit Model: The resulting closure is a compact algebraic equation. Unlike neural networks, it can be directly implemented in open-source solvers (like OpenLB) without external ML libraries, offering full transparency.
3. Zero-Shot Generalization: The model, trained exclusively on TGV and LDC flows, successfully generalizes to wall-bounded turbulent channel flow ($Re_\tau = 180$) without any supplemental wall-damping corrections or ad-hoc tuning.

4. Results and Validation

The discovered model was rigorously validated against DNS benchmarks and the standard Smagorinsky model across three regimes:

• Homogeneous Isotropic Turbulence (TGV):

  • The Φ-SO model aligned almost perfectly with DNS benchmarks for kinetic energy evolution and dissipation rates across all tested Reynolds numbers.
  • It accurately predicted the timing and amplitude of the peak dissipation rate, whereas the Smagorinsky model under-predicted this peak, indicating excessive damping during the transition phase.

• Inhomogeneous Shear Flow (LDC):

  • The model demonstrated superior performance in near-wall regions and high-shear boundary layers compared to Smagorinsky.
  • Crucially, it successfully captured secondary corner vortices (Moffatt eddies). The Smagorinsky model failed to sustain these delicate structures due to over-diffusion, while the symbolic model preserved the correct flow topology by dynamically adjusting viscosity in weak-shear regions.

• Generalization Test (Turbulent Channel Flow):

  • In a "blind test" on a turbulent channel flow ($Re_\tau = 180$) not present in the training data, the model reproduced the logarithmic law of the wall ($y^+ > 30$) with high fidelity.
  • It captured the peak of streamwise velocity fluctuations near the wall ($y^+ \approx 20$) significantly better than Smagorinsky, which tended to overestimate this peak.
  • The model maintained reasonable accuracy in the viscous sublayer without specific wall-damping functions.

5. Significance and Claims

The authors position this work as a paradigm shift from traditional derivation to data-driven discovery for turbulence modeling.

• Bridging the Gap: The study demonstrates that symbolic regression can bridge the gap between the high accuracy of data-driven methods and the interpretability of physical laws.
• Robustness: The model's ability to generalize to unseen wall-bounded flows without ad-hoc corrections suggests that the discovered algebraic structure inherently captures essential anisotropic mechanisms of turbulence, specifically the interaction between rotation and strain.
• Future Potential: While the specific coefficients and fractional exponents are currently empirical, the work highlights the potential of symbolic regression to uncover robust physical laws for the next generation of intelligent computational fluid dynamics (CFD) solvers.

Limitations Acknowledged by the Authors:

• The theoretical derivation linking the data-learned terms to first-principles turbulence theory remains complex and is an open topic.
• The model contains a ratio of invariants ($II_\Omega / II_S$) that could lead to numerical instability in regions where $II_S \to 0$ (e.g., pure solid-body rotation), requiring future regularization.
• Stability and accuracy in high-Reynolds-number engineering applications require further assessment.