arxiv: 2605.12304 · v1 · submitted 2026-05-12 · ⚛️ physics.flu-dyn

Recognition: no theorem link

Realizability-Constrained Machine Learning for Turbulence Closures in Wake Flows

Harshal D. Akolekar, Priyank H. Mehta, Talib Ansari

Pith reviewed 2026-05-13 03:33 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn

keywords turbulence closuresgene expression programmingrealizability constraintswake flowsCFD-driven machine learningsymbolic regressioncomputational fluid dynamicsmodel discovery

0 comments

The pith

Embedding barycentric realizability constraints into CFD-driven gene expression programming cuts turbulence model training cost by 42 percent while producing physically consistent wake closures.

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

The paper develops a method that inserts residual-based checks and a barycentric-map realizability filter directly inside the CFD solver loop during gene expression programming searches for turbulence models. This setup discards unstable or non-physical candidate expressions before they consume full simulation time. The result is a large drop in wasted effort and a near-elimination of non-realizable models at the end of training. A reader would care because conventional machine-learning turbulence closures frequently generate expressions that violate basic physical requirements and cause solver crashes or inaccurate wake predictions in engineering applications.

Core claim

By embedding two residual-based filtering criteria and a barycentric-map-based realizability constraint into the CFD solution loop of a gene expression programming framework, the method identifies and rejects invalid turbulence models early, yielding a 42.3 percent reduction in computational cost relative to the unfiltered baseline, a reduction of non-realizable models at convergence from 58.4 percent to 1.7 percent, and closures that improve mean wake predictions while remaining realizable and generalizable across a cylinder wake, a rectangular cylinder, an airfoil, and an axisymmetric body.

What carries the argument

The residual- and realizability-filtered CFD-driven gene expression programming framework, which applies barycentric-map realizability constraints and residual filters inside the CFD loop to reject invalid candidate models during training.

If this is right

The resulting turbulence models remain realizable across both training and unseen test cases.
Mean wake predictions improve while the models generalize to rectangular cylinders, airfoils, and axisymmetric bodies.
The framework supplies statistics on realizable model coefficients and the conditions that produce physically consistent wake behavior.
The same filtering approach supplies a scalable route to data-driven turbulence closures that satisfy stability and realizability from the outset.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same early-rejection logic could be ported to other symbolic-regression pipelines for turbulence modeling in boundary layers or jets.
By keeping nearly all final expressions physically admissible, the method may shorten the validation cycle needed before data-driven closures enter industrial CFD tools.
Testing whether the discovered closures retain accuracy at Reynolds numbers far above the training range would reveal how far the realizability filter preserves predictive power.

Load-bearing premise

Early rejection of non-realizable or unstable candidates via barycentric-map and residual filters does not unduly restrict the search space or bias the discovered models away from globally optimal closures for the target wake physics.

What would settle it

A demonstration that models discovered with the filters produce inaccurate mean wakes or non-realizable Reynolds stresses on a new geometry or Reynolds number outside the training distribution would show that the constraints have overly narrowed the search.

Figures

Figures reproduced from arXiv: 2605.12304 by Harshal D. Akolekar, Priyank H. Mehta, Talib Ansari.

**Figure 4.** Figure 4: The computational domain of the NACA 0012 [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

**Figure 6.** Figure 6: Filtered residual and realizability GEP-based CFD-driven framework. convergence-aware screening directly within the evolutionary loop. The key innovation is the integration of early-stage residual and realizability-based filtering mechanisms that terminate unproductive candidates before full convergence is attempted. Rather than treating solver stability as a posteriori information, it is incorporated as… view at source ↗

**Figure 7.** Figure 7: The barycentric map. A candidate clears the second stage only if Γ ≥ 𝛾min (10) or max𝑞 [𝑅(𝑞) (𝑁2 )] < 𝜖2 (10−6), where 𝛾min and 𝜖2 represents the minimum acceptable improvement factor (orderof-magnitude reduction) and convergence threshold. This condition ensures that residuals decrease at a sufficient rate and prevents stagnating models from consuming further computational resources. If the solution alr… view at source ↗

**Figure 8.** Figure 8: Comparison of computational speed for 25 generations of B-GEP, R-GEP and RR-GEP cases. 3. Results and Discussions This section evaluates the performance of the filtered CFD-driven framework in terms of computational efficiency, solver convergence behavior, cost function and model coefficient formulations. The predictive capability of the learned models is demonstrated through improved wake flow and reali… view at source ↗

**Figure 9.** Figure 9: Cost function evaluation over 25 generations. hours, and RR-GEP takes 13371 CPU hours, representing a 42.3% reduction in computational time in training. R-GEP reduces computational cost compared with the baseline BGEP by about 26.6%. This shows that realizability filtering is the more dominant factor in reducing computational time as it prevents unstable and non-realizable models from progressing to furt… view at source ↗

**Figure 10.** Figure 10: Residual plots of the best performing model in terms of cost function (a) B-GEP, (b) R-GEP, and (c) RR-GEP. (a) (b) (c) Ⅰ Ⅱ Ⅲ Ⅳ [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗

**Figure 11.** Figure 11: Barycentric map along three locations in the wake (I-III) and along the wake centreline (IV) for (a) B-GEP, (b) R-GEP, (c) RR-GEP. models, however, produce tightly clustered and well-aligned distributions across all wake locations, indicating bounded and consistent anisotropy evolution. Quantitatively, the variance perpendicular to the principal anisotropy trajectory is reduced by approximately an order … view at source ↗

**Figure 12.** Figure 12: Model rejection rate for R-GEP and RR-GEP with generation. to the number of generations. In the first generation, RGEP and RR-GEP reject 37.9% and 47.26% of the models. Across 50 generations, the average rejection rate for R-GEP is 21.5% and for RR-GEP is 37.8%. The model rejection rate trends thus highlight the role of filtering in stabilizing the evolutionary search. RR-GEP consistently rejects a large… view at source ↗

**Figure 14.** Figure 14: shows the magnitude of the mean values of |⟨𝜁1 ⟩| (𝜁1 is negative) and |⟨𝜁2 ⟩| ( 𝜁2 is positive) of five bestperforming models, at 𝑥∕𝐷 = 2.02 averaged from 𝑦∕𝐷 = 0 [PITH_FULL_IMAGE:figures/full_fig_p010_14.png] view at source ↗

**Figure 15.** Figure 15: Mean (a) 𝒂 𝐸𝐴𝑅𝑆𝑀 𝑋𝑋 and (b) 𝒂 𝐸𝐴𝑅𝑆𝑀 𝑋𝑌 for five best performing models at 𝑥∕𝐷 = 2.02 for y/D=0 to y/D = 0.91. dominated by streamwise normal stresses. The shear component, ⟨𝒂 EARSM 𝑋𝑌 ⟩, follows a similar trend and is the dominant contributor across all models. The RR-GEP formulation maintains lower shear stress levels leading to a more balanced representation of turbulence production and redistributio… view at source ↗

**Figure 17.** Figure 17: The separation bubble in the wake region of the cylinder at Re=3900 [PITH_FULL_IMAGE:figures/full_fig_p012_17.png] view at source ↗

**Figure 18.** Figure 18: (a) Wake profiles and (b) barycentric map for the rectangular 5:1 cylinder at 𝑥∕𝐻 = 4.5. turbulent diffusion, leading to an excessively deep and narrow wake with a minimum normalized velocity of approximately 0.1. In contrast, the RR-GEP model increases turbulent diffusion, resulting in a fuller and wider wake profile. The predicted minimum velocity increases to approximately 0.37, closely matching the… view at source ↗

**Figure 19.** Figure 19: (a) Wake profiles and (b) barycentric map for a NACA 0012 airfoil at 𝑥∕𝐶 = 1.05. realizability across fundamentally different flow regimes and Reynolds number which is three orders of magnitude higher. 4. Conclusions A residual- and realizability-filtered CFD-driven machine learning framework has been developed to address key limitations in turbulence model discovery using symbolic regression. The propo… view at source ↗

**Figure 20.** Figure 20: (a) Wake profiles and (b) barycentric map for the DARPA Suboff at 𝑥∕𝑊 = 12. improved anisotropy characteristics and physically consistent stress representations. Trained on a canonical cylinder wake at 𝑅𝑒 = 3900, the models improve wake prediction and also reduce separation bubble length from 𝑥∕𝐷 ≈ 4.5 to 2.25 and closely matches the LES reference dataset. Robust generalization is demonstrated across a r… view at source ↗

read the original abstract

Computational fluid dynamics (CFD)-driven machine learning frameworks based on symbolic regression offer a promising pathway for turbulence model discovery, but are often hindered by numerical instability, residual stagnation, and non-physical model behavior during training. In particular, realizability, which is rarely enforced explicitly during model development, remains a critical yet overlooked requirement, especially for accurate wake prediction. In this work, a residual- and realizability-filtered CFD-driven framework is proposed to enhance both efficiency and robustness within a gene expression programming (GEP) paradigm. The method integrates two residual-based filtering criteria along with a barycentric-map-based realizability constraint directly into the CFD solution loop, enabling early identification and rejection of unstable and non-realizable candidate models. This reduces unnecessary computational effort while guiding the search toward physically admissible solutions. The proposed approach achieves a 42.3% reduction in computational cost relative to the baseline CFD-driven GEP framework and reduces non-realizable models at convergence from 58.4% to 1.7%. The framework is trained on a canonical cylinder wake. The resulting models enhance mean wake prediction and remain realizable across training and test cases, with robust generalization to diverse geometries and operating conditions, including a rectangular cylinder, an airfoil, and an axisymmetric body. The study further provides insights into realizable model statistics, coefficient trends, and conditions governing physically consistent wake behavior. These results demonstrate that incorporating realizability and stability constraints within CFD-driven learning enables efficient and physically consistent turbulence model discovery, offering a scalable pathway toward reliable data-driven closure development.

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 realizability and residual filters speed up GEP training for wake turbulence closures and nearly eliminate bad models, but the accuracy improvements still need tighter verification.

read the letter

The filters make the GEP training for wake turbulence closures much cheaper and far more likely to produce realizable models. That is the key practical result from this paper. They integrate residual-based filters and a barycentric realizability constraint straight into the CFD solution loop during the evolutionary search. This allows early rejection of unstable or non-physical candidates. Compared to the baseline CFD-driven GEP, they report a 42.3 percent reduction in computational cost and a drop in non-realizable models at the end from 58.4 percent to 1.7 percent. The models trained on cylinder wake data then show better mean wake predictions and stay realizable when applied to a rectangular cylinder, an airfoil, and an axisymmetric body. The new element is this specific combination of filters inside the loop rather than after the fact. It guides the search toward admissible solutions without letting bad ones waste full CFD runs. The paper does a decent job showing generalization across geometries and operating conditions. That is useful for anyone who wants closures that work beyond the training case. They also give some statistics on realizable model behavior and coefficient trends. The soft spots are mostly around the strength of the accuracy claims. The abstract states improvements in mean wake prediction, but without error bars or detailed baseline comparisons it is hard to gauge the size of the gain. The early rejection approach could theoretically limit the search to models that are stable from the start, potentially missing better ones that take time to settle. The generalization results make this less of a problem, but a control experiment would have helped. This work is for fluid dynamicists and CFD practitioners interested in machine learning for turbulence modeling. Readers focused on practical implementation of symbolic regression will find the details on the filtering useful. It deserves a serious referee because the method is well-defined and the quantitative improvements are specific enough to check. I would send it to peer review.

Referee Report

3 major / 2 minor

Summary. The paper proposes a residual- and realizability-filtered CFD-driven gene expression programming (GEP) framework for discovering turbulence closures in wake flows. It integrates barycentric-map realizability constraints and residual-based filters into the CFD loop to reject unstable or non-realizable candidates early, claiming a 42.3% reduction in computational cost relative to baseline CFD-driven GEP, a drop in non-realizable models at convergence from 58.4% to 1.7%, improved mean wake predictions on a canonical cylinder, and robust generalization to rectangular cylinder, airfoil, and axisymmetric body cases while maintaining realizability.

Significance. If the central claims hold, the work offers a practical route to more efficient and physically consistent turbulence model discovery by embedding stability and realizability enforcement directly in the search loop. The reported efficiency gains and cross-geometry generalization would be valuable for wake-dominated flows where standard closures often fail, provided the discovered models are not merely the first admissible ones but demonstrably competitive with or superior to unconstrained optima.

major comments (3)

[Method and Results (filter integration and model statistics)] The load-bearing assumption that early rejection via barycentric-map realizability and residual filters preserves access to globally optimal realizable closures (rather than truncating the search to the first admissible candidates) is not adequately tested. The manuscript should compare the wake-prediction accuracy and generalization performance of models evolved with versus without the early-rejection filters; without this, the 42.3% cost reduction and 1.7% non-realizable fraction could simply reflect aggressive pruning rather than discovery of superior physics.
[Results (quantitative performance claims)] Quantitative gains are reported without error bars, multiple random seeds, or explicit baseline implementation details (e.g., population size, CFD convergence criteria, or exact definition of the 58.4% baseline non-realizable rate). This weakens the claim that the filtered framework “enhances mean wake prediction” beyond the baseline; a table or figure showing mean and standard deviation of drag/lift coefficients or wake deficit errors across repeated runs is needed.
[Generalization tests] Verification that the discovered closures satisfy the Navier-Stokes equations beyond mean wake statistics is limited to the training geometry. The paper should report residual norms or higher-order statistics (e.g., Reynolds-stress anisotropy) on the test geometries to confirm that realizability enforcement translates into consistent momentum balance rather than merely admissible but inaccurate closures.

minor comments (2)

[Method] Notation for the barycentric-map realizability constraint should be defined explicitly (e.g., the mapping from Reynolds-stress invariants to the barycentric triangle) rather than referenced only by name.
[Results] The abstract and results mention “insights into realizable model statistics, coefficient trends” but no corresponding figure or table is referenced; adding one would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's thorough review and valuable suggestions for improving the manuscript. We address each of the major comments below, providing clarifications and indicating the revisions we plan to implement.

read point-by-point responses

Referee: [Method and Results (filter integration and model statistics)] The load-bearing assumption that early rejection via barycentric-map realizability and residual filters preserves access to globally optimal realizable closures (rather than truncating the search to the first admissible candidates) is not adequately tested. The manuscript should compare the wake-prediction accuracy and generalization performance of models evolved with versus without the early-rejection filters; without this, the 42.3% cost reduction and 1.7% non-realizable fraction could simply reflect aggressive pruning rather than discovery of superior physics.

Authors: We agree that a direct comparison between the filtered and unfiltered frameworks is necessary to substantiate that the early-rejection mechanism does not unduly restrict the search space to suboptimal solutions. In the revised manuscript, we will include additional experiments comparing the wake-prediction accuracy and generalization performance of models evolved with and without the filters. This will demonstrate that the proposed approach not only reduces computational cost but also maintains or improves model quality by focusing on physically admissible candidates. revision: yes
Referee: [Results (quantitative performance claims)] Quantitative gains are reported without error bars, multiple random seeds, or explicit baseline implementation details (e.g., population size, CFD convergence criteria, or exact definition of the 58.4% baseline non-realizable rate). This weakens the claim that the filtered framework “enhances mean wake prediction” beyond the baseline; a table or figure showing mean and standard deviation of drag/lift coefficients or wake deficit errors across repeated runs is needed.

Authors: We acknowledge that the quantitative claims would benefit from statistical validation. We will provide explicit details on the baseline implementation, including population size, CFD convergence criteria, and the definition of the non-realizable rate. Additionally, we will conduct multiple runs with different random seeds and report mean values with standard deviations for key metrics such as drag and lift coefficients and wake deficit errors in a new table or figure. revision: yes
Referee: [Generalization tests] Verification that the discovered closures satisfy the Navier-Stokes equations beyond mean wake statistics is limited to the training geometry. The paper should report residual norms or higher-order statistics (e.g., Reynolds-stress anisotropy) on the test geometries to confirm that realizability enforcement translates into consistent momentum balance rather than merely admissible but inaccurate closures.

Authors: We thank the referee for highlighting the importance of verifying consistency beyond mean statistics. In the revised version, we will report residual norms of the Navier-Stokes equations and higher-order statistics, including Reynolds-stress anisotropy, for the discovered models on the test geometries (rectangular cylinder, airfoil, and axisymmetric body). This will confirm that the realizability enforcement leads to consistent momentum balance across cases. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical performance metrics from constrained search

full rationale

The paper describes an algorithmic CFD-driven GEP framework augmented with barycentric realizability and residual filters. All quantitative claims (42.3% cost reduction, drop from 58.4% to 1.7% non-realizable models) are measured outcomes of executing the constrained evolutionary search on cylinder-wake training data and testing on other geometries. No derivation chain, fitted parameter renamed as prediction, or self-citation load-bearing uniqueness theorem is present; the method is self-contained against external CFD benchmarks and does not reduce any result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on the domain assumption that realizability (via barycentric map) and residual stability are the dominant barriers to successful GEP turbulence discovery, plus standard GEP evolutionary operators and CFD solver numerics. No new physical entities are introduced.

axioms (2)

domain assumption Realizability of Reynolds stresses is a necessary condition for physical turbulence models in incompressible flows
Invoked to justify the barycentric-map filter as a hard constraint during model evolution.
domain assumption Gene expression programming can evolve closure expressions that satisfy the filtered Navier-Stokes equations when instability is removed early
Underlies the claim that early rejection improves both efficiency and final model quality.

pith-pipeline@v0.9.0 · 5592 in / 1381 out tokens · 74066 ms · 2026-05-13T03:33:48.101436+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

[1]

R. D. Sandberg, V. Michelassi, Fluid Dynamics of Axial Turboma- chinery: Blade- and Stage-Level Simulations and Models, Annual Review of Fluid Mechanics 54 (2022) 255–285

work page 2022
[2]

G. Schmitt, About Boussinesq’s Turbulent Viscosity Hypothesis: Historical Remarks and a Direct Evaluation of its Validity., Comptes Rendus Mecanique, Elsevier Masson 335 (2007) 617–627

work page 2007
[3]

M.Leschziner,StatisticalTurbulenceModellingforFluidDynamics- Demystified:anIntroductoryTextforGraduateEngineeringStudents, World Scientific, Oxford, UK, 2015

work page 2015
[4]

S. B. Pope, A More General Effective-Viscosity Hypothesis, Journal of Fluid Mechanics 72 (1975) 331–340

work page 1975
[5]

J. Ling, J. Templeton, A. Kurzawski, Reynolds Averaged Turbulence Modelling Using Deep Neural Networks with Embedded Invariance, J. Fluid Mech 807 (2016) 155–166

work page 2016
[6]

Weatheritt, R

J. Weatheritt, R. D. Sandberg, A Novel Evolutionary Algorithm Applied to Algebraic Modifications of the RANS Stress-Strain Re- lationship, Journal of Computational Physics 325 (2016) 22–37

work page 2016
[7]

M.L.Kaandorp,R.P.Dwight,Data-drivenmodellingoftheReynolds stress tensor using random forests with invariance, Computers and Fluids 202 (2020) 104497

work page 2020
[8]

Schmelzer, R

M. Schmelzer, R. P. Dwight, P. Cinnella, Discovery of Algebraic Reynolds-Stress Models Using Sparse Symbolic Regression, Flow, Turbulence and Combustion 104 (2020) 579–603

work page 2020
[9]

Reissmann, Y

M. Reissmann, Y. Fang, A. S. H. Ooi, R. D. Sandberg, Symbolic tur- bulence model development for complex-geometry flows exploiting language model-based transfer learning, Physics of Fluids 37 (2025) 095180

work page 2025
[10]

Weatheritt, R

J. Weatheritt, R. D. Sandberg, The development of algebraic stress models using a novel evolutionary algorithm, International Journal of Heat and Fluid Flow 68 (2017) 298–318

work page 2017
[11]

H. D. Akolekar, J. Weatheritt, N. Hutchins, R. D. Sandberg, G. Laskowski, V. Michelassi, Development and Use of Machine- LearntAlgebraicReynoldsStressModelsforEnhancedPredictionof Wake Mixing in Low-Pressure Turbines, J. Turbomach. 141 (2019)

work page 2019
[12]

H. D. Akolekar, R. D. Sandberg, N. Hutchins, V. Michelassi, G.Laskowski,Machine-LearntTurbulenceClosuresforLowPressure Turbineswith UnsteadyInflowConditions, Journalof Turbomachin- ery 141 (2019) 101009

work page 2019
[13]

H. D. Akolekar, F. Waschkowski, Y. Zhao, R. Pacciani, R. D. Sand- berg, Transition Modeling for Low Pressure Turbines Using CFD- Driven Machine Learning, Energies 14 (2021) 4680

work page 2021
[14]

Y. Gu, Y. Fang, H. Akolekar, R. Pacciani, M. Marconcini, A. Ooi, R. Sandberg, Machine-learning strategies for transition/turbulence modelling for low pressure turbines with unsteady inflow conditions, in: ISUAAAT, Melbourne, Australia, 2025, p. 1

work page 2025
[15]

Weatheritt, R

J. Weatheritt, R. Pichler, R. D. Sandberg, G. Laskowski, V.Michelassi,MachineLearningforTurbulenceModelDevelopment Using a High Fidelity HPT Cascade Simulation, in: ASME Turbo Expo 2017: Turbomachinery Technical Conference and Exposition, Charlotte, 2017, pp. 2017–63497

work page 2017
[16]

Y.Zhao,H.D.Akolekar,J.Weatheritt,V.Michelassi,R.D.Sandberg, RANS Turbulence Model Development Using CFD-Driven Machine Learning, Journal of Computational Physics 411 (2020) 109413

work page 2020
[17]

H.D.Akolekar,Y.Zhao,R.D.Sandberg,N.Hutchins,V.Michelassi, Turbulence Model Development for Low & High Pressure Turbines Using a Machine Learning Approach, in: International Society for Air Breathing Engines, Canberra, Australia, 2019, p. 1

work page 2019
[18]

Schoepplein, J

M. Schoepplein, J. Weatheritt, R. Sandberg, M. Talei, M. Klein, Ap- plication of an evolutionary algorithm to LES modelling of turbulent transport in premixed flames, Journal of Computational Physics 374 (2018) 1166–1179

work page 2018
[19]

Haghiri, R

A. Haghiri, R. D. Sandberg, Large Eddy Simulations of Wall Jets With Coflow for the Study of Turbulent Prandtl Number Variations and Data-driven Modeling, Phys. Rev. Fluids 5 (2020) 64501

work page 2020
[20]

R. D. Sandberg, R. Tan, J. Weatheritt, A. Ooi, A. Haghiri, V. Michelassi, G. Laskowski, Applying Machine Learnt Explicit Algebraic Stress and Scalar Flux Models To a Fundamental Trailing Edge Slot, Journal of Turbomachinery 140 (2018) 101008

work page 2018
[21]

Parneix, D

S. Parneix, D. Laurence, P. Durbin, A procedure for using DNS databases, ASME. J. Fluids Eng. 120 (1998) 40–47

work page 1998
[22]

Waschkowski, Y

F. Waschkowski, Y. Zhao, R. Sandberg, J. Klewicki, Multi-objective CFD-driven development of coupled turbulence closure models, J. Comput. Phys 452 (2022) 110922

work page 2022
[23]

X. Xu, F. Waschkowski, A. S. Ooi, R. D. Sandberg, Towards robust and accurate Reynolds-averaged closures for natural convection via multi-objective CFD-driven machine learning, Int. J. Heat Mass Transf. 187 (2022) 122557

work page 2022
[24]

H. D. Akolekar, F. Waschkowski, R. D. Sandberg, R. Pacciani, Y. Zhao, Integrated Transition and Wake Mixing Predictions in Low Pressure, in:ProceedingsoftheASMETurboExpo2022,Rotterdam, The Netherlands, 2022, pp. 1–13

work page 2022
[25]

Y. Fang, Y. Zhao, F. Waschkowski, A. S. Ooi, R. D. Sandberg, To- wardMoreGeneralTurbulenceModelsviaMulticaseComputational- Fluid-Dynamics-Driven Training, AIAA Journal 61 (2023) 2100– 2115

work page 2023
[26]

Marconcini, A Data-Driven Approach for Generalizing the LaminarKineticEnergyModelforSeparationandBypassTransition inLow-andHigh-PressureTurbines, ASMEJTurbomach146(2024) 091005

Y.Fang,Y.Zhao,H.D.Akolekar,A.S.Ooi,R.D.Sandberg,R.Pac- ciani, M. Marconcini, A Data-Driven Approach for Generalizing the LaminarKineticEnergyModelforSeparationandBypassTransition inLow-andHigh-PressureTurbines, ASMEJTurbomach146(2024) 091005

work page 2024
[27]

Y. Fang, M. Reissmann, R. Pacciani, Y. Zhao, A. S. Ooi, M. Mar- concini, H. D. Akolekar, R. D. Sandberg, Accelerating cfd-driven trainingof transitionandturbulencemodels forturbineflowsby one- shotandreal-timetransformerintegration, ComputersandFluids306 (2026) 106927

work page 2026
[28]

J. L. Lumley, Toward a turbulent constitutive relation, Journal of Fluid Mechanics 41 (1970) 413

work page 1970
[29]

S.Banerjee,R.Krahl,F.Durst,C.Zenger, Presentationofanisotropy properties of turbulence, invariants versus eigenvalue approaches, Journal of Turbulence 8 (2007) 1–27

work page 2007
[30]

thesis, University of Melbourne, 2019

H.D.Akolekar,TurbulenceModelDevelopmentandImplementation For Low Pressure Turbines Using a Machine Learning Approach, Ph.D. thesis, University of Melbourne, 2019

work page 2019
[31]

J.-X.Wang,J.-L.Wu,H.Xiao,APhysicsInformedMachineLearning Approach for Reconstructing Reynolds Stress Modeling Discrepan- cies Based on DNS Data, Physical Review Fluids 2 (2017) 034603

work page 2017
[32]

J. G. Wnek, C. Schrock, E. M. Wolf, M. Wolff, A polar turbulence invariantmapwithapplicabilitytorealisablemachinelearningturbu- lence models, Journal of Fluid Mechanics 1018 (2025) A36

work page 2025
[33]

Parnaudeau, J

P. Parnaudeau, J. Carlier, D. Heitz, E. Lamballais, Experimental and numerical studies of the flow over a circular cylinder at reynolds number 3900, Physics of Fluids 20 (2008) 085101. Ansari et al.:Preprint submitted to ElsevierPage 15 of 16 Realizability-Constrained Machine Learning for Turbulence Closures in Wake Flows

work page 2008
[34]

Mannini, A

C. Mannini, A. Mariotti, L. Siconolfi, M. V. Salvetti, Benchmark on the aerodynamics of a 5:1 rectangular cylinder: Further experimental and les results, in: Proceedings of the 11th Workshop on DLES, Springer International Publishing, 2019, pp. 427–432

work page 2019
[35]

Y. Lin, M. Savill, N. R. Vadlamani, R. Jefferson-Loveday, Wall- resolved large eddy simulation over naca0012 airfoil, International Journal of Aerospace Sciences 2 (2013) 149–162

work page 2013
[36]

J. M. Jimenez, M. Hultmark, A. J. Smits, The intermediate wake of a body of revolution at high reynolds numbers, Journal of Fluid Mechanics 659 (2010) 516–539

work page 2010
[37]

F. R. Menter, Two-equation eddy-viscosity turbulence models for engineering applications, AIAA Journal 32 (1994) 1598–1605

work page 1994
[38]

FMFP2025–INS–112

T.Ansari,P.H.Mehta,H.D.Akolekar, Cfd-drivenmachinelearning for improving wake prediction for cylinders, in: Proceedings of the 12thInternationaland52ndNationalFMFPConference,Ahmedabad, India, 2025, pp. FMFP2025–INS–112

work page 2025
[39]

S.V.Patankar,NumericalHeatTransferandFluidFlow,Hemisphere Publishing Corporation, Washington, DC, 1980

work page 1980
[40]

H. K. Versteeg, W. Malalasekera, An Introduction to Computational Fluid Dynamics: The Finite Volume Method, 2 ed., Pearson Educa- tion, 2007

work page 2007
[41]

J.H.Ferziger,M.Perić,ComputationalMethodsforFluidDynamics, Springer, 2020

work page 2020
[42]

R. A. Horn, C. R. Johnson, Matrix Analysis, Cambridge University Press, 2013

work page 2013
[43]

McConkey, N

R. McConkey, N. Kalia, E. Yee, F.-S. Lien, Realisability-informed machine learning for turbulence anisotropy mappings, Journal of Fluid Mechanics 1019 (2025) A49. Ansari et al.:Preprint submitted to ElsevierPage 16 of 16

work page 2025