Field Inversion Symbolic Regression with Embedded Equation Learner for Interpretable Turbulence Model Correction

This paper proposes FISR-EQL, an interpretable framework that embeds equation learning into a PDE-constrained field inversion process to derive compact, physics-consistent analytical corrections for turbulence models, achieving performance comparable to neural networks while significantly improving separated flow predictions without sacrificing transparency.

The Gist

The Big Picture: Fixing the "Weather Forecast" of Airflow

Imagine you are trying to predict the weather. You have a supercomputer running a complex model, but it keeps getting the rain wrong in specific areas. It's not that the computer is slow; it's that the rules (the physics equations) inside the computer are slightly too simple to handle the messy reality of the wind.

In the world of airplanes, this is a huge problem. Engineers use simulations (called RANS) to design wings and engines. These simulations are fast, but they often get separation wrong.

• Separation is when the air stops sticking to the wing and starts swirling chaotically behind it.
• If the simulation thinks the air stays stuck when it actually separates, the plane might stall (lose lift) unexpectedly.

For decades, scientists tried to fix this by tweaking numbers (parameters) in the equations. But sometimes, the problem isn't the numbers; it's the structure of the equation itself.

The Old Way: The "Two-Step" Mistake

Previously, scientists used a "Two-Step" approach to fix these models:

1. Step 1 (The Detective): They ran a simulation, compared it to real data (like wind tunnel results), and figured out where the model was wrong. They created a "correction map" for that specific flight.
2. Step 2 (The Translator): They took that messy correction map and tried to teach a "Black Box" AI (a Neural Network) to guess the pattern.

The Problem: This is like a detective writing a report, then hiring a translator who doesn't speak the language perfectly. The translation often loses meaning.

• Black Box: The AI gives a result, but no one knows why or how it got there.
• Objective Mismatch: The AI learns to match the map, not the physics. When you put that AI back into the flight simulator, it often crashes or gives weird results because it wasn't trained to play by the rules of physics directly.

The New Way: FISR-EQL (The "Smart Architect")

This paper introduces a new method called FISR-EQL. Think of it as hiring a Smart Architect who builds the correction directly into the blueprint, rather than adding a patch later.

Here is how it works, broken down into simple concepts:

1. The "Equation Learner" (EQL)

Instead of using a Black Box AI that spits out numbers, the researchers use a special type of AI called an Equation Learner.

• The Analogy: Imagine a standard AI is a chef who tastes a soup and says, "Add 3.42 grams of salt." You have no idea why.
• The EQL Chef: This chef says, "Add salt if the soup is too sour, but only if it's hot." The EQL learns mathematical formulas (like $y = x^2 + 5$) instead of just numbers.
• The Result: The final output is a clear, readable math equation that engineers can actually understand and trust.

2. The "End-to-End" Training

In the old way, the detective and the translator worked separately. In this new way, they work together in one room.

• The Analogy: Imagine training a pilot. In the old way, you taught them to fly a simulator, then taught them to read a manual. In the new way, you put the manual inside their brain and train them to fly the plane while reading the manual simultaneously.
• Why it matters: The model learns to fix the airflow while obeying the laws of physics. It doesn't just guess; it optimizes the math directly to match reality.

3. The "Shield" (Protecting the Good Parts)

The researchers knew that their new correction was great for messy, swirling air (separation), but it might mess up smooth, calm air (attached flow).

• The Analogy: Think of the correction as a sunscreen. You want to apply it to your face (the turbulent areas) to protect you, but you don't want to smear it all over your body (the smooth airflow), or you'll get a rash.
• They built a "shield" into the math that automatically turns the correction OFF when the air is smooth and ON when the air is swirling.

What Did They Find?

They tested this new "Smart Architect" on two famous tricky shapes: a curved step and a hump on a wall.

• The Result: The new model fixed the swirling air problems much better than the old models.
• The Surprise: It performed almost as well as the "Black Box" AI (which is usually the most accurate), but with a huge bonus: It is transparent. We can look at the final equation and say, "Ah, I see! It adds extra viscosity when the rotation is high."

Does it Work on New Things? (Generalization)

The real test was: "If we teach it on a step and a hump, will it work on a totally different airplane wing?"

• They tested it on a Periodic Hill (a bumpy road for air), a Cube in the wind, and a complex High-Lift Airfoil (a wing with flaps).
• The Outcome: It worked beautifully. It predicted when the plane would stall (lose lift) much more accurately than before, without ruining the performance of the smooth parts of the wing.
• Why? Because the model learned the physics (the "why"), not just the shape (the "what").

Summary: Why This Matters

This paper presents a bridge between Data Science and Physics.

• Old Way: Fast but opaque (Black Box) or accurate but messy (Two-Step).
• New Way (FISR-EQL): Fast, accurate, and interpretable.

It gives engineers a tool that is as smart as a super-computer AI but as understandable as a high-school physics textbook. This means we can design safer, more efficient airplanes with confidence, knowing exactly why the computer made its predictions.

Technical Summary

1. Problem Statement

Traditional Reynolds-Averaged Navier-Stokes (RANS) turbulence models, such as the Shear-Stress-Transport (SST) model, suffer from "model inadequacy" due to simplified physical assumptions. This is particularly evident in complex flow regimes involving separation, where RANS models often underestimate Reynolds stress, leading to an overprediction of separation bubble sizes and poor reattachment predictions.

While data-driven approaches like Field Inversion and Machine Learning (FIML) have been proposed to correct these errors, they face significant limitations:

• Two-Stage Approaches (FIML-classic/FISR): These separate the process into inferring a spatial correction field and then fitting a model to it. This introduces objective mismatch, where the fitted model does not directly minimize the error in the governing equations, often leading to accumulated errors and numerical instability when deployed.
• Neural Network-Based End-to-End (FIML-direct): While these optimize the model directly within the PDE constraints, they rely on "black-box" neural networks. This lacks interpretability and often exhibits weaker generalization capabilities compared to symbolic expressions.

The core challenge is to develop a turbulence correction framework that is end-to-end optimal (minimizing the PDE-constrained error directly), interpretable (yielding explicit analytical expressions), and generalizable across different flow configurations.

2. Methodology: FISR-EQL Framework

The authors propose FISR-EQL (Field Inversion Symbolic Regression with Embedded Equation Learner), a novel framework that embeds an Equation Learner (EQL) network directly into the field inversion process.

• Core Architecture:

  • Instead of a standard neural network with uniform activation functions (e.g., ReLU), the EQL uses a heterogeneous set of symbolic operators (unary: $id, \tanh, 1/(1+|x|)$; binary: multiplication).
  • The network takes local flow features (e.g., production-to-dissipation ratio $P/\epsilon$, vortex-based Reynolds number $Re_\omega$, tensor invariants $\lambda_2, \lambda_5$) as inputs and outputs a correction factor $\beta$.
  • This $\beta$ modifies the production term of the turbulent kinetic energy ($k$) in the SST model.

• End-to-End Optimization:

  • The framework treats the EQL network parameters as design variables within a PDE-constrained optimization problem.
  • Using the discrete adjoint method (via the DAFoam solver), gradients of the objective function (mismatch between RANS and reference data) are computed directly with respect to the network parameters.
  • This eliminates the intermediate step of inferring a correction field, ensuring the final model directly minimizes the flow field error.

• Sparsity and Interpretability:

  • To ensure the resulting expression is compact and interpretable, the training process employs a three-phase strategy:
    1. Phase 1: Standard training with data misfit and $L_2$ regularization on the correction field.
    2. Phase 2: Introduction of $L_1$ regularization on network weights to enforce sparsity (driving weak connections to zero).
    3. Phase 3: Pruning of negligible branches and fine-tuning of the remaining parameters.
  • This process yields a sparse weight matrix from which a concise explicit analytical expression can be extracted.

• Shielding Mechanism:

  • A shielding function ($f_s$) is applied to ensure the correction is only active in separated shear layers and not in attached boundary layers, preserving the accuracy of skin-friction predictions.

3. Key Contributions

1. Unified Optimization: The paper introduces the first framework that combines the optimality of end-to-end FIML-direct with the interpretability of symbolic regression. It solves the "objective mismatch" problem inherent in two-stage methods.
2. Embedded Equation Learner: By replacing standard activation functions with symbolic operators within the adjoint-based inversion loop, the method enables the direct discovery of analytical correction terms.
3. Robust Generalization: The method demonstrates that a single, compact analytical expression derived from two training cases can generalize effectively to unseen, complex 3D and high-lift configurations without retraining.
4. Practical Deployability: Unlike black-box neural networks, the resulting explicit formulas are easy to implement in standard CFD solvers, ensuring stability and physical consistency.

4. Results

The framework was applied to the SST model and trained on two canonical separated flows: the Curved Backward-Facing Step (CBFS) and the NASA Hump.

• Training Performance:

  • The selected sparse expression achieved a reattachment location error of 0.28% (Hump) and 32.99% (CBFS) relative to the baseline SST error.
  • While FIML-direct (neural network) achieved slightly lower error on training data due to higher parameter capacity, FISR-EQL provided a clear analytical form.
  • The derived expression (e.g., involving terms like $\lambda_2$ and $Re_\omega$) physically correlates with shear layer characteristics, confirming interpretability.

• Generalization (Test Cases):

  • Periodic Hills: The model reduced the Root-Mean-Square Error (RMSE) of velocity profiles by 12% to 31% across different slope parameters, significantly outperforming the baseline SST.
  • NLR7301 High-Lift Airfoil: The model successfully delayed flow separation, accurately predicting the stall angle and maximum lift coefficient, which the baseline SST underestimated.
  • FAITH Hill (3D Separation): The model reduced mean velocity errors by 29.67% compared to SST. Notably, neural-network-based baselines failed to activate corrections in this unseen 3D geometry, whereas the explicit expression remained robust.
  • Zero-Pressure-Gradient (ZPG) Flat Plate: The shielding function successfully prevented the correction from degrading attached boundary layer predictions, maintaining skin-friction accuracy identical to the baseline SST.

5. Significance

This work represents a significant step forward in interpretable machine learning for fluid dynamics. It bridges the gap between data-driven accuracy and physical transparency. By providing a pathway to derive compact, analytical turbulence corrections that are optimized directly against governing equations, FISR-EQL offers a practical solution for improving RANS simulations in industrial aerodynamic design. It addresses the critical need for models that are not only accurate but also understandable, stable, and generalizable across diverse flow regimes, paving the way for future applications in unsteady flow problems.