A few-shot and physically restorable symbolic regression turbulence model based on normalized general effective-viscosity hypothesis

This paper proposes a few-shot, physically restorable symbolic regression turbulence model based on a normalized general effective-viscosity hypothesis that achieves high accuracy and generalizability across diverse flow regimes using limited training data.

The Gist

The "Smart Chef" Approach to Predicting Turbulent Chaos

Imagine you are trying to predict how a massive, swirling storm will move, or how smoke curls from a cigarette. In physics, this is called turbulence. It is one of the most chaotic and difficult things in the universe to predict.

Engineers usually use a "shortcut" called RANS. Think of RANS as a standard recipe book for a chef. It’s fast and works okay for basic meals (like boiling an egg), but if you ask the chef to make a complex, five-course molecular gastronomy feast, the recipe book fails. The food comes out wrong because the "shortcut" isn't smart enough to handle the complexity.

Recently, scientists tried using AI (Machine Learning) to fix this. But AI has a problem: it’s a "data glutton." To learn how to cook, an AI needs to see millions of different recipes and ingredients. In fluid dynamics, getting that much data is incredibly expensive and slow—it's like needing to eat a billion meals just to learn how to season a steak.

This paper introduces a smarter way: The "Few-Shot, Physically Restorable" Model.

---

1. The "Few-Shot" Learner: The Intuitive Chef

Instead of an AI that needs to see a billion meals, the authors created a model that uses "Few-Shot Learning."

The Analogy: Imagine a chef who has only ever practiced making simple pasta dishes (this is the "limited training data"). However, because they understand the logic of heat, salt, and texture, when you suddenly ask them to cook a complex spicy curry, they don't panic. They use their fundamental understanding to adapt.

The researchers trained their model only on one specific type of flow (periodic hills). Even though the model had never "seen" a jet engine or an airplane wing, it was able to predict how they would behave because it learned the underlying rules of the dance, not just the moves.

2. Symbolic Regression: The "Math Poet"

Most AI is a "Black Box"—it gives you an answer, but it can't tell you why. It’s like a chef who makes a great soup but can't tell you the ingredients.

The authors used something called Symbolic Regression. Instead of just giving a number, this method searches for a mathematical formula. It’s like a "Math Poet" that looks at the chaos and writes a beautiful, short poem (an equation) that explains exactly what is happening. Because we have an actual formula, humans can read it, understand it, and trust it.

3. "Physically Restorable": The Safety Net

One big fear with AI in engineering is that it might "hallucinate" or do something physically impossible.

The Analogy: Imagine a self-driving car. You want it to be high-tech and smart, but if it enters a simple, straight highway, you want it to behave like a standard, predictable car. You don't want it trying to do "fancy" maneuvers when a simple one works perfectly.

The authors designed their model to be "Physically Restorable." This means that in simple situations (like water flowing smoothly over a flat plate), the AI "steps back" and reverts to the old, reliable, standard recipe (the baseline model). It only uses its "fancy AI powers" when the flow gets truly wild and complicated. This ensures the model is both cutting-edge and safe.

---

The Result: A Universal Translator for Chaos

The researchers tested their "Intuitive Chef" on everything from airplane wings to high-speed jet engine rotors.

• The Old Way: Was either too slow (perfect accuracy) or too inaccurate (the shortcut).
• The Standard AI Way: Was too hungry for data and hard to trust.
• This New Way: Learned from a tiny bit of data, wrote down its own rules in math, and knew exactly when to be "smart" and when to be "simple."

In short: They taught a computer to understand the "grammar" of turbulence, so it can write its own "stories" about how fluids move, even in worlds it has never visited before.

Technical Summary

Technical Summary: A Few-Shot and Physically Restorable Symbolic Regression Turbulence Model

1. Problem Statement

The Reynolds-averaged Navier-Stokes (RANS) method is the industry standard for rapid engineering turbulence simulations due to its low computational cost. However, its accuracy is fundamentally limited by the empirical nature of classical turbulence models (e.g., $k$-$\omega$-SST). While recent data-driven approaches (Machine Learning/Deep Learning) have attempted to improve accuracy, they face three critical bottlenecks:

• High Computational Cost: Training on massive, multi-case datasets is expensive and slows down the development cycle.
• Data Scarcity: High-fidelity data (DNS/LES) for complex industrial flows is prohibitively difficult to acquire.
• Lack of Generalizability and Physical Consistency: Many models fail when applied to flow regimes different from their training data, and they often lack the ability to "revert" to proven classical models in regimes where those models are already accurate (e.g., the viscous sublayer).

2. Methodology

The authors propose a novel framework based on the normalized general effective-viscosity hypothesis and Symbolic Regression (SR). The approach is characterized by two unique concepts:

• Few-Shot Learning Paradigm: Unlike traditional ML that requires massive datasets, this model is trained on a very limited set of flow configurations (specifically, 3D slices of DNS periodic hill flows) but is tested on substantially different physics (airfoils, compressors, flat plates).
• Physically Restorable Symbolic Regression: Instead of a "black-box" neural network, the authors use the PySR library to evolve explicit mathematical expressions for the tensor basis coefficients ($\hat{g}_i$). By using a normalized tensor basis, the model is designed to automatically "vanish" or revert to the baseline Boussinesq hypothesis in specific physical regimes (e.g., when certain dimensionless numbers approach limits).

Technical Workflow:

1. Feature Engineering: The model uses tensor invariants ($I_1 \dots I_{17}$) and extra features ($q_1 \dots q_5$) such as the Q-criterion, wall-distance-based Reynolds number, and ratios of turbulent to mean strain time scales.
2. Normalization: All tensors and features are normalized to ensure the relative influence of each term can be assessed by the magnitude of its coefficient.
3. Symbolic Training: The SR algorithm searches for the simplest mathematical functions that map the input features to the normalized Reynolds stress anisotropy coefficients derived from DNS data.
4. Implementation: The resulting explicit equations are embedded into a CFD solver (TCAE/OpenFOAM) for RANS simulations.

3. Key Contributions

• Formalization of "Few-Shot" Turbulence Modeling: Defining a metric where a model is judged by its ability to generalize from a narrow subset of physics to a broad spectrum of turbulent flows.
• Explicit Mathematical Discovery: Moving away from neural networks toward interpretable, symbolic equations that describe the Reynolds stress anisotropy.
• Automatic Physical Reversion: Demonstrating that a model trained on one type of flow (periodic hills) can autonomously recover classical Boussinesq behavior in near-wall regions of a flat plate without external "shielding" functions.

4. Results

The authors compared two models: SR 3T (3-tensor model for 2D flows) and SR 5T (5-tensor model).

• Periodic Hill Flows: The SR 5T model significantly outperformed the baseline $k$-$\omega$-SST and the SR 3T model in predicting separation zones and velocity profiles.
• Zero Pressure Gradient Flat Plate: The SR 5T model successfully maintained the accuracy of the baseline model. It demonstrated "physical restorability" by reverting to the classical model in the viscous sublayer ($y^+ < 2$) where the additional tensor terms became inactive.
• NACA0012 Airfoil: The SR 5T model provided superior accuracy in predicting the lift coefficient, particularly in the pre-stall and high-angle-of-attack regimes.
• NASA Rotor 37 (Transonic Compressor): Despite being trained only on incompressible data, the SR 5T model accurately captured shock-wave characteristics and significantly improved the prediction of adiabatic efficiency and total pressure ratio compared to the baseline. It even outperformed a neural network model (SA-NN) that had been trained directly on the compressor's experimental data.

5. Significance

This work represents a significant shift in data-driven turbulence modeling. It proves that interpretability and physical constraints are more important than data volume. By using symbolic regression, the authors have created a model that is:

1. Efficient: Requires minimal training data.
2. Robust: Generalizes across different Mach numbers (incompressible to transonic) and geometries.
3. Reliable: Does not "break" the physics in well-understood regimes (like the near-wall region), making it a viable candidate for integration into industrial CFD workflows.