An Initial Attempt of Converged Machine-Learning Assisted Turbulence Modeling in RANS Simulations with Eddy-Viscosity Hypothesis
Pith reviewed 2026-05-25 01:45 UTC · model grok-4.3
The pith
Machine learning framework lets RANS simulations converge to DNS mean velocity and viscosity profiles in channel flows.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The MLATM framework learns target turbulence quantities from high-fidelity data using non-dimensional variables, incorporates a prior estimate from conventional models, and closes the loop between the machine-learning predictor and the RANS solver so that the computed mean flow and eddy viscosity converge to the DNS reference in channel-flow tests.
What carries the argument
The closed-loop computational chain that repeatedly inserts machine-learning predictions of turbulence quantities into the RANS solver until the solution converges to DNS mean profiles.
If this is right
- Mean velocity profiles from RANS can be made to match DNS data in the tested channel flows.
- Turbulence viscosity profiles can also be recovered consistently with the DNS training set.
- The closed loop prevents divergence that would otherwise arise from direct substitution of learned quantities into the RANS equations.
- The eddy-viscosity hypothesis remains intact while the model coefficients are supplied by data.
Where Pith is reading between the lines
- The same non-dimensional variable set might allow the framework to be retrained on other simple shear flows where DNS data exist.
- If the loop is retained, the method could supply data-driven corrections inside existing industrial RANS codes without changing their convergence properties.
- Extension to flows with separation would require new non-dimensional inputs that capture the additional physics while preserving the closed-loop stability.
Load-bearing premise
That reasonable non-dimensional variables can be chosen so the machine-learning predictions steer the RANS solver back to the DNS mean flow field under the eddy-viscosity hypothesis.
What would settle it
Apply the trained MLATM model to a different geometry such as a backward-facing step and check whether the predicted mean velocity profile deviates from independent DNS data.
read the original abstract
This work presents a converged framework of Machine-Learning Assisted Turbulence Modeling (MLATM). Our objective is to develop a turbulence model directly learning from high fidelity data (DNS/LES) with eddy-viscosity hypothesis induced. First, the target machine-learning quantity is discussed in order to avoid the ill-conditioning problem of RANS equations. Then, the novel framework to build the turbulence model using the prior estimation of traditional models is demonstrated. A close-loop computational chain is designed to ensure the convergence of result. Besides, reasonable non-dimensional variables are selected to predict the target learning variables and make the solver re-converge to DNS mean flow field. The MLATM is tested in incompressible turbulent channel flows, and it proved that the result converges well to DNS training data for both mean velocity and turbulence viscosity profiles.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript presents an initial framework called Machine-Learning Assisted Turbulence Modeling (MLATM) for RANS simulations under the eddy-viscosity hypothesis. It selects target quantities to avoid ill-conditioning, incorporates prior traditional model estimates, implements a closed-loop chain to enforce convergence, chooses non-dimensional input variables, and tests the approach on incompressible turbulent channel flows, claiming that the solver re-converges to DNS mean velocity and turbulence viscosity profiles.
Significance. If the central convergence claim holds under scrutiny and the variable selection proves robust, the closed-loop design could address a practical barrier in data-driven RANS modeling by ensuring the learned eddy viscosity remains consistent with the discretized momentum equations. The explicit focus on avoiding ill-conditioning and using DNS training data is a constructive step, though the absence of architectural or metric details limits evaluation of whether the result generalizes beyond the specific training cases.
major comments (2)
- [Abstract] Abstract: the assertion that 'reasonable non-dimensional variables are selected to predict the target learning variables and make the solver re-converge to DNS mean flow field' is load-bearing for the central claim yet is stated without derivation, invariance analysis, sensitivity study, or comparison to alternatives; this leaves open whether the inputs are case-specific or consistent with the Boussinesq hypothesis across flows.
- [Abstract] Abstract: the statement that 'it proved that the result converges well to DNS training data for both mean velocity and turbulence viscosity profiles' is unsupported by any quantitative error metrics, residual norms, grid-convergence checks, or description of the ML architecture, training procedure, or data partitioning; without these the convergence result cannot be verified as framework-driven rather than data-specific fitting.
Simulated Author's Rebuttal
We thank the referee for the constructive comments on the abstract and the overall framework. We address each major comment below and agree that the abstract requires expansion to better support the central claims. The revised manuscript will incorporate additional details on variable selection and quantitative convergence metrics.
read point-by-point responses
-
Referee: [Abstract] Abstract: the assertion that 'reasonable non-dimensional variables are selected to predict the target learning variables and make the solver re-converge to DNS mean flow field' is load-bearing for the central claim yet is stated without derivation, invariance analysis, sensitivity study, or comparison to alternatives; this leaves open whether the inputs are case-specific or consistent with the Boussinesq hypothesis across flows.
Authors: The non-dimensional inputs were selected in the manuscript to respect the eddy-viscosity hypothesis and to ensure invariance under Galilean transformations and scaling consistent with the Boussinesq assumption. However, the abstract is indeed too terse and omits the physical rationale and any sensitivity checks. We will revise the abstract to briefly state the selection criteria and add a short subsection (or expand the existing discussion) that includes invariance arguments, a sensitivity study on the chosen variables, and explicit comparison to alternative non-dimensional sets. This will clarify that the choice is not purely case-specific. revision: yes
-
Referee: [Abstract] Abstract: the statement that 'it proved that the result converges well to DNS training data for both mean velocity and turbulence viscosity profiles' is unsupported by any quantitative error metrics, residual norms, grid-convergence checks, or description of the ML architecture, training procedure, or data partitioning; without these the convergence result cannot be verified as framework-driven rather than data-specific fitting.
Authors: The current manuscript demonstrates convergence primarily through profile comparisons in the results section and relies on the closed-loop design to enforce consistency with the discretized RANS equations. We agree that the abstract claim would be stronger with quantitative support. In the revision we will (i) report L2-norm errors and maximum deviations for both mean velocity and eddy viscosity, (ii) describe the neural-network architecture, training procedure, and train/test partitioning, and (iii) confirm that the underlying RANS solver was grid-converged independently of the ML model. These additions will be placed in the methods and results sections and referenced from the abstract. revision: yes
Circularity Check
ML fit to DNS data reproduces DNS profiles by construction in closed-loop RANS
specific steps
-
fitted input called prediction
[Abstract]
"Besides, reasonable non-dimensional variables are selected to predict the target learning variables and make the solver re-converge to DNS mean flow field. The MLATM is tested in incompressible turbulent channel flows, and it proved that the result converges well to DNS training data for both mean velocity and turbulence viscosity profiles."
The ML model is fitted directly to the DNS training data for the identical channel-flow case; the non-dimensional inputs and target quantity are chosen specifically so that the RANS solver re-converges to those same DNS profiles. The reported convergence is therefore the expected output of the fitting process under the eddy-viscosity closure, not an independent test of the framework.
full rationale
The framework trains an ML model on DNS data for channel flow to predict a target quantity (eddy viscosity) under the eddy-viscosity hypothesis, then uses a closed-loop RANS solver that is shown to re-converge to the identical DNS mean flow. The abstract explicitly states that non-dimensional variables and the target quantity are selected to enable this re-convergence. Because the training data and test case are the same, the reported match to DNS velocity and viscosity profiles follows directly from the supervised fit rather than from an independent derivation or out-of-sample validation. No external benchmark or first-principles derivation of the input variables is quoted.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Eddy-viscosity hypothesis is valid for the incompressible channel flows considered.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.