Correcting Turbulence Models with AI – and Understanding Why

by

Author: Denis Avetisyan

A new framework combines data-driven optimization with symbolic regression to refine turbulence simulations while maintaining interpretability.

This work introduces FISR-EQL, a method for learning interpretable corrections to existing turbulence models via field inversion and symbolic regression, achieving performance comparable to neural networks.

Despite advances in computational fluid dynamics, turbulence modeling remains a critical bottleneck in accurately predicting complex flow phenomena. This challenge is addressed in ‘Field Inversion Symbolic Regression with Embedded Equation Learner for Interpretable Turbulence Model Correction’, which introduces FISR-EQL, a novel framework that directly optimizes interpretable analytical corrections to existing turbulence models within a PDE-constrained field inversion process. By embedding symbolic regression and equation learning, FISR-EQL achieves performance comparable to data-driven neural networks while retaining full transparency and physical consistency. Could this approach unlock a new paradigm for developing optimal, yet understandable, turbulence models for engineering applications?

The Inherent Limitations of Empirical Turbulence Models

Engineering simulations routinely employ turbulence models, such as the Shear Stress Transport (SST) model, to approximate the chaotic behavior of fluids. However, these models inherently rely on simplifications of the complex interactions within turbulent flows, often leading to inaccuracies when predicting real-world phenomena. This limitation becomes particularly pronounced in scenarios involving flow separation, swirling motions, or strong pressure gradients, where the assumptions baked into the model break down. Consequently, designs optimized using such simulations may exhibit discrepancies between predicted and actual performance, necessitating costly adjustments and potentially compromising safety or efficiency. The inability to accurately capture turbulence, therefore, represents a significant bottleneck in fields reliant on computational fluid dynamics, driving ongoing research into more robust and reliable modeling techniques.

The limitations of current turbulence models arise from fundamental simplifications made to represent the chaotic nature of fluid motion. These models often rely on assumptions about flow behavior that hold true only under specific conditions, hindering their ability to accurately predict turbulence across diverse scenarios. A particularly challenging case involves separated flows – where the fluid stream detaches from a surface – because these flows exhibit strong three-dimensional effects and complex interactions that are poorly captured by standard models. Consequently, predictions in these regimes can deviate significantly from experimental observations, impacting the reliability of simulations used in engineering design and scientific analysis. The inability to adapt to varying flow conditions, especially in complex scenarios like separated flows, represents a significant hurdle in achieving truly accurate and robust turbulence modeling.

Advancements in accurately simulating turbulent flows are fundamentally linked to progress across a spectrum of engineering disciplines. In aerodynamics, refined turbulence modeling promises more efficient wing designs and reduced drag, leading to fuel savings and enhanced aircraft performance. Similarly, within hydrodynamics, precise simulations are essential for optimizing ship hulls, improving propeller efficiency, and predicting ocean currents with greater accuracy. Perhaps most critically, combustion processes-central to power generation, internal combustion engines, and even climate modeling-are profoundly sensitive to turbulence; improved fidelity in these simulations enables the design of cleaner, more efficient engines and a more complete understanding of atmospheric chemistry. Ultimately, elevating the accuracy of turbulence models isn’t merely an academic pursuit, but a practical necessity for innovation and optimization in fields that shape modern technology and address pressing global challenges.

Recent advancements in computational fluid dynamics leverage data-driven techniques like Field Inversion and Machine Learning (FIML) to refine turbulence models, offering a potential solution to the limitations of traditional approaches. These methods aim to correct model deficiencies by learning directly from high-fidelity data, such as Direct Numerical Simulation or experimental measurements. However, a significant challenge remains in ensuring the reliability and understanding of these corrected models; FIML often produces ‘black box’ corrections, where the underlying physical principles guiding the improvement are obscured. This lack of interpretability hinders the ability to extrapolate the corrected model to novel flow conditions and raises concerns about its robustness – meaning the model’s performance may degrade unexpectedly outside the specific dataset used for training. Further research focuses on developing FIML techniques that balance accuracy with transparency, allowing engineers to confidently deploy these advanced models in critical applications.

FISR-EQL: A Framework for Deriving Analytical Corrections

FISR-EQL employs Equation Learner (EQL), a symbolic regression technique implemented via a neural network, within a PDE-constrained optimization loop. EQL functions by searching for analytical equations that best fit provided data, effectively learning relationships between variables. This learned equation is then incorporated as a correction term within the Partial Differential Equation (PDE) solver. The optimization process adjusts the parameters of the learned equation to minimize a defined loss function, subject to the constraints imposed by the governing PDE. This integration allows FISR-EQL to directly discover and implement analytical corrections, differing from methods that rely on parameterized functional mappings. The process leverages the data-driven nature of neural networks with the transparency of symbolic expressions, resulting in a framework capable of learning interpretable model improvements.

FISR-EQL leverages Equation Learner (EQL) to generate analytical expressions representing corrections to existing turbulence models. Unlike black-box approaches, this yields corrections formulated as mathematical equations – for example, \Delta k = f(x, y, z, k, \epsilon) – directly relating corrected variables to relevant flow features. This explicit formulation improves interpretability, allowing users to understand how and why the model is being adjusted. Furthermore, expressing corrections analytically enforces physical consistency; the resulting equations are subject to dimensional analysis and can be assessed for adherence to known physical constraints, mitigating the risk of unphysical predictions often associated with purely data-driven model learning.

The Adjoint Method is employed within FISR-EQL to calculate the gradients of the cost function with respect to the turbulence model parameters during optimization. This technique involves solving the adjoint equation, a modified form of the governing equations, to efficiently determine the sensitivity of the cost function to changes in these parameters. Compared to finite difference or direct differentiation methods, the Adjoint Method significantly reduces computational cost, particularly for high-dimensional parameter spaces, as it avoids repeated forward simulations for each parameter variation. The gradient information obtained is then used to update the model parameters via an optimization algorithm, iteratively minimizing the discrepancy between simulations and target data.

Traditional Field Inversion and Machine Learning (FIML) methods often struggle with generalizing learned corrections beyond the training dataset and ensuring physical consistency due to the black-box nature of the machine learning component. FISR-EQL addresses these limitations by integrating Equation Learner (EQL) directly within the field inversion loop; this allows the framework to optimize for analytical expressions rather than arbitrary parameter sets. Specifically, the iterative process alternates between solving the governing equations with a turbulence model incorporating the EQL-discovered correction term, and updating the EQL-generated expression based on the resulting discrepancies. This embedded approach constrains the learned correction to a symbolic form, promoting generalization capability and ensuring the correction adheres to known physical principles, unlike the purely data-driven nature of standard FIML implementations.

Validation Across Established Separated Flow Test Cases

FISR-EQL underwent rigorous testing utilizing established canonical separated flow scenarios – the NASA Hump Case, the CBFS Case, and the Surface-Mounted Cube Case – to assess its predictive capabilities. Results demonstrate a quantifiable improvement in the accurate determination of separation and reattachment points within these flows. Specifically, the framework’s ability to identify the location of these critical flow features surpassed that of baseline models across all three test cases, indicating a heightened fidelity in representing the physics of separated flow phenomena. These canonical cases provide a standardized means of comparison and validation against existing computational fluid dynamics methodologies.

Validation of the FISR-EQL framework was conducted using the NLR7301 Airfoil Case to assess its performance in predicting aerodynamic stall. Results indicate a substantial improvement in stall prediction accuracy when compared to baseline models. Specifically, the framework demonstrated an enhanced capability to accurately identify the angle of attack at which stall occurs, leading to more reliable predictions of lift and drag characteristics near the stall point. This improved stall prediction is critical for applications requiring accurate aerodynamic modeling, such as aircraft design and performance analysis.

Evaluation of the FISR-EQL framework on the Periodic Hill Case demonstrated its ability to accurately predict flow behavior in geometrically complex scenarios. This assessment utilized a standardized test case to quantify performance, resulting in a reduction in Root Mean Squared Error (RMSE) of up to 31.84% when compared against baseline models. This improvement indicates a substantial gain in predictive capability for flows influenced by repeating geometric features, highlighting the framework’s generalization performance beyond simpler flow conditions.

FISR-EQL demonstrates maintained accuracy when applied to attached flow scenarios, confirmed by validation against the Zero-Pressure-Gradient Flat Plate test case. Further assessment using the FAITH Hill test case reveals improvements in velocity profile prediction, achieving up to a 29.67% reduction in error compared to baseline models. This indicates the framework does not sacrifice performance on simpler flow regimes while enhancing its capabilities in more complex separated flow conditions.

Towards a Paradigm of Transparent and Robust Fluid Dynamics

A critical challenge in refining Reynolds-Averaged Navier-Stokes (RANS) turbulence models lies in avoiding spurious corrections that destabilize simulations, particularly within boundary layers. The Fluid Interface Shielding with Reynolds-based Error-correction, Quadratic Linearization (FISR-EQL) framework addresses this through a dedicated Shielding Function. This function acts as a safeguard, effectively diminishing model corrections in regions where the flow remains attached and well-behaved – specifically, within the viscous sublayer and near-wall regions. By selectively applying corrections only where they are demonstrably needed-in regions of separation or strong adverse pressure gradients-FISR-EQL prevents the introduction of unphysical behavior and ensures the preservation of accurate near-wall dynamics, ultimately leading to more stable and reliable turbulence simulations.

The framework known as FISR-EQL distinguishes itself by not simply applying corrections to fluid dynamics simulations, but by deriving those corrections as analytical expressions. This means the adjustments made to the model aren’t opaque, empirically-determined values, but rather, are rooted in mathematical relationships that can be inspected and understood. Consequently, researchers gain access to the reasoning behind each correction, fostering confidence in the simulation’s results and enabling a clear assessment of potential uncertainties. This level of transparency is crucial for building trustworthy predictive models, as it allows for rigorous validation and facilitates the identification of areas where the underlying physics might be incompletely captured – ultimately accelerating progress towards more reliable and insightful fluid dynamics simulations.

The development of an interpretable fluid dynamics framework offers more than just improved simulation accuracy; it provides a pathway to fundamentally deeper understanding of turbulent flow. Unlike ‘black box’ models, this approach yields analytical expressions for corrections, allowing researchers to dissect why a simulation produces a particular result, rather than simply accepting that it does. This transparency is critical for validating physical assumptions embedded within the model and for identifying previously unknown relationships governing complex flows. Consequently, the ability to interrogate the simulation process fosters hypothesis generation and targeted experimentation, dramatically accelerating the pace of scientific discovery in areas ranging from aerodynamic design to weather prediction and even biomedical engineering. The framework empowers researchers to move beyond merely predicting fluid behavior to truly unraveling the underlying physics.

The current framework, while demonstrating success in fundamental flow scenarios, is poised for expansion into more challenging and realistic simulations. Researchers are actively developing strategies to apply FISR-EQL – the framework for robust and interpretable fluid dynamics – to geometries beyond simple configurations, including those with intricate curves, varying cross-sections, and complex surface textures. Simultaneously, efforts are underway to extend its capabilities to encompass a wider range of flow regimes, such as those characterized by high Reynolds numbers, strong pressure gradients, or the presence of multiple interacting turbulent structures. Successful implementation across these diverse conditions promises a significant leap towards truly predictive fluid dynamics, enabling accurate simulations that can be confidently used for design optimization, performance forecasting, and a deeper understanding of complex physical phenomena.

The pursuit of accurate turbulence modeling, as detailed in this work, echoes a fundamental principle of mathematical consistency. The FISR-EQL framework, by directly optimizing analytical expressions, embodies this pursuit. It prioritizes a provable, interpretable correction-a solution grounded in established equations-rather than relying on the ‘black box’ nature of neural networks. As Galileo Galilei observed, “You cannot teach a man anything; you can only help him discover it himself.” This framework doesn’t simply predict corrections; it allows for the discovery of underlying relationships within the turbulence models, aligning with the principle that true understanding stems from demonstrable, mathematical truths, not merely observed performance.

What Remains Constant?

The pursuit of turbulence model correction, as exemplified by FISR-EQL, frequently resembles a sophisticated parameter fitting exercise. One achieves demonstrable improvement, yet the underlying question persists: does the correction truly reflect a deeper understanding of the physics, or merely a skillful interpolation within the training data? Let N approach infinity – what remains invariant? The framework’s strength lies in its insistence on analytical expression, a commendable resistance to the black-box tendencies prevalent in contemporary machine learning. However, the limitations of symbolic regression itself – the combinatorial explosion of possible expressions – cannot be ignored. Future work must address the scalability of this approach to more complex flows and potentially explore hybrid methods that strategically combine the interpretability of analytical forms with the representational power of neural networks.

A critical, often overlooked, aspect is the validation process. Demonstrating performance comparable to neural networks is a necessary, but insufficient, condition for success. The true test lies in extrapolation – applying the corrected model to regimes significantly different from those used for training. A rigorous investigation of generalization capabilities, coupled with sensitivity analysis to identify the most influential terms in the correction, is paramount. The framework’s reliance on adjoint methods, while efficient, introduces computational cost; exploration of alternative optimization strategies may prove fruitful.

Ultimately, the field must confront a fundamental tension: the desire for both accuracy and interpretability. One cannot simply declare a model ‘corrected’ based on reduced error metrics. The corrected terms should ideally possess a clear physical meaning, providing insights into the deficiencies of the original model. The path forward likely involves a more synergistic integration of data-driven techniques with first-principles reasoning, striving not merely to predict, but to understand.

Original article: https://arxiv.org/pdf/2604.14569.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-19 09:38