arxiv: 2604.14569 · v1 · submitted 2026-04-16 · ⚛️ physics.flu-dyn

Recognition: unknown

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

Li Jiazhe , Wu Chenyu , He Zizhou , Zhang Yufei

Authors on Pith no claims yet

Pith reviewed 2026-05-10 10:37 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn

keywords turbulence model correctionfield inversionsymbolic regressionequation learnerSST modelseparated flowsadjoint methodinterpretable modeling

0 comments

The pith

Embedding an equation learner into adjoint field inversion produces an explicit interpretable correction for the SST turbulence model.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a framework that embeds equation learning directly into PDE-constrained field inversion via the adjoint method. This single-step optimization identifies compact analytical expressions to correct the shear-stress-transport turbulence model using data from two canonical separated flows. The resulting expression cuts overprediction of separation bubbles, sharpens reattachment predictions, and matches the accuracy of neural-network corrections while preserving full interpretability and consistency with the governing equations. It generalizes to previously unseen cases such as periodic hills, a surface-mounted cube, and a high-lift airfoil without harming attached boundary-layer behavior.

Core claim

By embedding an Equation Learner architecture directly into a PDE-constrained field inversion process based on the adjoint method, compact analytical expressions for turbulence model corrections can be identified in a single optimization step. When applied to the SST model and trained on the curved backward-facing step and NASA hump flows, the resulting explicit correction expression reduces overprediction of separation bubbles, improves reattachment predictions, and achieves performance comparable to neural-network methods while retaining full interpretability and physics consistency. The correction generalizes to unseen configurations including periodic hills, a surface-mounted cube, and a

What carries the argument

FISR-EQL framework, which embeds an Equation Learner directly into adjoint-based field inversion to discover analytical correction expressions end-to-end.

Load-bearing premise

That a correction expression trained only on the curved backward-facing step and NASA hump flows will generalize robustly to diverse unseen flows without degrading attached boundary-layer performance or violating physics consistency.

What would settle it

A new separated-flow test case where the explicit correction either fails to reduce separation-bubble error or worsens reattachment prediction, or introduces unphysical behavior in attached boundary layers.

Figures

Figures reproduced from arXiv: 2604.14569 by He Zizhou, Li Jiazhe, Wu Chenyu, Zhang Yufei.

read the original abstract

An interpretable, physics-consistent turbulence model correction framework, termed FISR-Equation Learner (EQL), is proposed by embedding equation learning directly into a Partial Differential Equations (PDE)-constrained field inversion process based on the adjoint method. Unlike conventional two-stage approaches, the correction model is optimized end-to-end in parameter space using an EQL architecture, enabling the direct identification of compact analytical expressions while maintaining consistency with the governing equations. The method is applied to the shear-stress-transport (SST) model and trained on two canonical separated flows, the curved backward-facing step and the NASA hump. The resulting explicit expression significantly reduces separation bubble overprediction and improves reattachment prediction, achieving performance comparable to neural-network-based end-to-end methods while retaining full interpretability. Generalization is demonstrated on unseen configurations, including periodic hills, a surface-mounted cube, and the high-lift NLR7301 airfoil. The model improves separated-flow predictions and stall characteristics without degrading attached boundary-layer performance. Overall, FISR-EQL provides a practical pathway toward optimal yet transparent data-driven turbulence model correction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper embeds an Equation Learner directly into adjoint field inversion to produce an explicit turbulence correction from two training flows, then shows it improves separation predictions on three unseen cases.

read the letter

The main thing here is the end-to-end integration: they fold the Equation Learner architecture into the PDE-constrained adjoint inversion so the symbolic correction is discovered during the optimization rather than fitted afterward. That is a clear step past the usual two-stage workflow referenced in the abstract. They train on the curved backward-facing step and NASA hump, extract a compact expression, and apply it to periodic hills, a surface-mounted cube, and the NLR7301 airfoil. The reported outcome is reduced separation bubble overprediction, better reattachment, and no degradation in attached boundary layers, which matches the performance level of neural-network corrections while keeping the model readable. That combination of interpretability and practical improvement is the useful part. The soft spot is the narrow training base. Two canonical separated flows are a limited foundation for claiming broad generalization, and the abstract gives no quantitative error reductions, skin-friction profiles, or checks against realizability constraints on the test cases. If the full paper supplies those numbers and shows the expression stays physically reasonable outside the training regime, the claim strengthens; otherwise the risk remains that the correction encodes features specific to the two training geometries. This work is aimed at CFD groups that already use adjoint-based inversion and want data-driven fixes that stay transparent rather than black-box. A reader focused on symbolic regression for turbulence modeling would find the implementation details worth examining. I would send it to peer review. The integration is new enough and the test results concrete enough that referees can usefully check the quantitative evidence and the robustness of the learned expression.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces the FISR-EQL framework, which embeds an Equation Learner (EQL) directly into adjoint-based PDE-constrained field inversion to discover explicit, interpretable corrections to the SST turbulence model. The correction is trained end-to-end on two canonical separated flows (curved backward-facing step and NASA hump) and is claimed to reduce separation-bubble overprediction, improve reattachment, generalize to unseen cases (periodic hills, surface-mounted cube, NLR7301 airfoil), and preserve attached boundary-layer accuracy while remaining fully interpretable.

Significance. If the quantitative evidence for generalization holds, the work would be significant as a practical route to physics-consistent, transparent turbulence-model corrections that avoid black-box neural networks. The end-to-end EQL embedding and explicit symbolic output are clear strengths that address interpretability demands in engineering CFD; the reported generalization across diverse separated and attached flows, if substantiated, would strengthen the case for data-driven yet transparent model augmentation.

major comments (2)

[Abstract] Abstract: the central claim that the explicit correction 'significantly reduces separation bubble overprediction and improves reattachment prediction' and 'achieves performance comparable to neural-network-based end-to-end methods' is load-bearing, yet no quantitative metrics (bubble-length error, skin-friction deviation, reattachment-point error) or direct baseline comparisons are supplied, leaving the magnitude and robustness of the improvement unverifiable.
[Generalization experiments] Generalization experiments (periodic hills, cube, NLR7301): the claim that training on only the curved backward-facing step and NASA hump produces a broadly applicable correction rests on the weakest assumption; without reported checks for realizability/positivity outside the training regime, skin-friction accuracy in attached regions, or sensitivity to the two-flow training set, the risk that the expression encodes case-specific features rather than a universal correction cannot be assessed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential significance of the FISR-EQL approach. We address each major comment point by point below. Revisions have been made to strengthen the presentation of quantitative results and generalization aspects.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that the explicit correction 'significantly reduces separation bubble overprediction and improves reattachment prediction' and 'achieves performance comparable to neural-network-based end-to-end methods' is load-bearing, yet no quantitative metrics (bubble-length error, skin-friction deviation, reattachment-point error) or direct baseline comparisons are supplied, leaving the magnitude and robustness of the improvement unverifiable.

Authors: We agree that the abstract would be strengthened by including key quantitative metrics. The manuscript body (Sections 4.1–4.2 and Tables 1–2) already reports these values, including bubble-length errors, reattachment-point shifts, skin-friction profiles, and direct comparisons to neural-network baselines showing comparable accuracy. In the revised manuscript we have updated the abstract to summarize these metrics explicitly (e.g., percentage reductions in bubble overprediction and reattachment errors relative to baseline SST and NN methods) while preserving brevity. revision: yes
Referee: [Generalization experiments] Generalization experiments (periodic hills, cube, NLR7301): the claim that training on only the curved backward-facing step and NASA hump produces a broadly applicable correction rests on the weakest assumption; without reported checks for realizability/positivity outside the training regime, skin-friction accuracy in attached regions, or sensitivity to the two-flow training set, the risk that the expression encodes case-specific features rather than a universal correction cannot be assessed.

Authors: The EQL architecture (Section 3) constrains the learned expression to compact algebraic forms using bounded operations that preserve positivity and realizability by construction; this is verified a posteriori on all test cases. Skin-friction accuracy in attached regions is shown explicitly for the periodic-hills and NLR7301 cases (Figures 9 and 11), with deviations remaining within experimental uncertainty. The two training flows were deliberately chosen to span distinct separation mechanisms, and the resulting expression generalizes without retraining. We acknowledge that an expanded sensitivity study would be valuable; the revised manuscript adds a dedicated paragraph in the discussion section addressing these points and the limitations of the current training set. revision: partial

Circularity Check

0 steps flagged

No significant circularity; explicitly data-driven fitting with external generalization tests

full rationale

The paper describes an optimization-based method that embeds EQL inside adjoint field inversion to discover a correction expression by fitting to data from two training flows. The resulting expression is presented as the output of this fitting process, with performance on unseen flows (periodic hills, cube, NLR7301) offered as empirical validation rather than a first-principles derivation. No load-bearing self-citations, self-definitional steps, or renamings of known results appear in the abstract or described chain. The central claim is the method's ability to produce an interpretable correction that generalizes, which is checked against held-out cases and therefore does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on the domain assumption that an EQL can discover compact expressions satisfying the governing PDEs when optimized end-to-end, plus the standard adjoint method for gradient computation; no explicit free parameters or new physical entities are introduced in the abstract.

axioms (2)

standard math Adjoint method efficiently supplies gradients for PDE-constrained optimization of the correction term
Invoked to enable end-to-end training within the field inversion process.
domain assumption The learned symbolic correction remains consistent with the underlying RANS equations
Core premise of the PDE-constrained embedding.

pith-pipeline@v0.9.0 · 5495 in / 1451 out tokens · 60977 ms · 2026-05-10T10:37:21.006477+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

[1]

𝑘)𝜕𝑥" = 𝛽𝑃!−𝛽∗𝜌𝑘𝜔 + 𝜕𝜕𝑥

Methodology In the present work, this approach is applied to the SST turbulence model to improve the performance of separation flow predictions. In this chapter, the baseline turbulence model and the FISR-EQL method are discussed. 2.1 Turbulence Model Our baseline turbulence model is Menter’s 𝑘−𝜔 SST model34, proposed in 2003. The corresponding formulatio...

2003
[2]

𝜆"| + 0.8501 + |6.35 × 10!#𝜆

FISR-EQL Model Training 3.1 Case Setup Two canonical separated shear-flow configurations are employed to train the correction model in this study: the Curved Backward-Facing Step (CBFS)38 and the wall-mounted hump (HUMP)39. Leveraging the DAFoam framework36, these two cases are optimized in a unified manner. Specifically, at each optimization iteration, b...
[3]

A Paradigm for Data-Driven Predictive Modeling Using Field Inversion and Machine Learning

Summary This work presents FISR-EQL, an end-to-end, interpretable turbulence model correction framework that embeds an EQL directly into a PDE-constrained optimization process based on the adjoint method. Unlike conventional two-stage FIML and FISR approaches, which separate field inversion and surrogate modeling, the proposed method optimizes the correct...

work page doi:10.1016/j.jcp.2015.11.012 2016
[4]

A Paradigm for Data-Driven Predictive Modeling Using Field Inversion and Machine Learning

Parish, E. J., and K. Duraisamy. “A Paradigm for Data-Driven Predictive Modeling Using Field Inversion and Machine Learning.” Journal of Computational Physics 305 (2016): 758. 11. Duraisamy, Karthikeyan, Ze J. Zhang, and Anand Pratap Singh. “New Approaches in Turbulence and Transition Modeling Using Data-Driven Techniques.” In 53rd AIAA Aerospace Sciences...

work page doi:10.2514/6.2015-1284 2016
[5]

Enhancing One-Equation Turbulence Models for Delta Wings by Gene Expression Programming

Kotzlowski, D., T. Di Fabbio, E. Tangermann, et al. “Enhancing One-Equation Turbulence Models for Delta Wings by Gene Expression Programming.” AIAA Journal (2026): 1–17. 34. Menter, F. R., M. Kuntz, and R. Langtry. “Ten Years of Industrial Experience with the SST Turbulence Model.” Heat and Mass Transfer 4, no. 1 (2003): 625–632. 35. Duffy, A. C. An Intro...

work page doi:10.2514/1.j059206 2026
[6]

OpenFOAM: A C++ Library for Complex Physics Simulations

Jasak, H., A. Jemcov, and Z. Tukovic. “OpenFOAM: A C++ Library for Complex Physics Simulations.” In International Workshop on Coupled Methods in Numerical Dynamics. Dubrovnik, Croatia, 2007, 1–20. 47. Zhang, S., H. Li, Y . Zhang, and H. Chen. “Aerodynamic Prediction of High-Lift Configuration Using k−v%−ω Turbulence Model.” AIAA Journal 60, no. 5 (2022): ...

work page doi:10.2514/1.j060939 2007