pith. sign in

arxiv: 2605.26358 · v2 · pith:OQI72EDJnew · submitted 2026-05-25 · ⚛️ physics.flu-dyn · cs.LG

Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows

Daniel Dehtyriov , Jonathan F. MacArt , Justin Sirignano This is my paper

Pith reviewed 2026-06-29 20:05 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn cs.LG
keywords Reynolds-averaged Navier-Stokesturbulence modelingalgebraic Reynolds stress modeldeep learning closuremachine learning for CFDsquare duct flowperiodic hill flowweak equilibrium assumption
0
0 comments X

The pith

A neural network embedded in an algebraic Reynolds stress equation improves RANS velocity predictions by 2-4 times on average and generalizes from attached to separated flows without retraining.

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

The paper introduces the Deep Algebraic Reynolds Stress Model (DARSM), which embeds a neural network inside an implicit algebraic closure derived from the Reynolds stress transport equations under the weak-equilibrium assumption. The network learns mappings from flow invariants to closure parameters, and the entire system is optimized end-to-end through the governing PDEs using derived adjoint equations that respect the solver's implicit-explicit structure. On square-duct and periodic-hill test cases, DARSM lowers average velocity error relative to standard RANS by factors of 2-4 across Reynolds numbers and geometries, with some cases improving by 12 times. The model trained only on attached, anisotropy-dominated duct flows transfers directly to separated hill flows, a change in underlying physics regime. It also outperforms five other machine-learning closure approaches that either train offline or bypass the equations entirely.

Core claim

DARSM trains a neural network to supply empirical coefficients inside an implicit algebraic Reynolds stress relation obtained from the full transport equations under the weak-equilibrium assumption; end-to-end differentiation through the coupled RANS solver via custom adjoint equations removes distribution shift, yielding 2-4 times lower average velocity error than baseline RANS on both square-duct and periodic-hill benchmarks while generalizing from attached to separated regimes without retraining.

What carries the argument

The Deep Algebraic Reynolds Stress Model (DARSM), a neural network that supplies parameters to an implicit algebraic Reynolds stress equation derived under the weak-equilibrium assumption, optimized through the governing PDEs with adjoint equations that exploit the solver's implicit-explicit structure.

If this is right

  • RANS simulations of attached and mildly separated turbulent flows can achieve 2-4 times lower average velocity error with a single trained DARSM model.
  • A model trained exclusively on square-duct data transfers directly to periodic-hill flows without retraining, indicating regime-level generalization.
  • End-to-end optimization through the coupled implicit closure eliminates the distribution shift that affects offline-trained ML closures.
  • DARSM outperforms offline training, tensor-basis networks, field-inversion methods, DeepONets, and physics-informed networks on the same benchmarks.
  • The approach requires only small high-fidelity datasets because the algebraic structure supplies most of the physics.

Where Pith is reading between the lines

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

  • The same hybrid structure could be tested on other RANS closures that admit algebraic reductions, such as scalar-flux or heat-transfer models.
  • If the weak-equilibrium assumption holds across a wider class of engineering flows, DARSM-style models might reduce the need for full Reynolds-stress transport simulations in many industrial cases.
  • The adjoint derivation technique may apply to other stiff implicit-explicit solvers in computational fluid dynamics where direct differentiation fails.

Load-bearing premise

The weak-equilibrium assumption used to collapse the Reynolds stress transport equations into an implicit algebraic form stays accurate enough for the attached duct flows in training and the separated hill flows in testing.

What would settle it

Measure whether DARSM still reduces velocity error by at least 2 times relative to baseline RANS when applied to a new separated flow geometry whose mean-flow statistics visibly violate the weak-equilibrium assumption.

Figures

Figures reproduced from arXiv: 2605.26358 by Daniel Dehtyriov, Jonathan F. MacArt, Justin Sirignano.

Figure 1
Figure 1. Figure 1: Hybrid discrete adjoint pipeline. Top row: the three outer steps (forward solve, iterative solve for λ, gradient extraction). The matvec (∂R/∂U) ⊤v is assembled by applying M forward sub-steps in reverse and applying the three-step recipe (bottom box) at each substep: each substep reuses the efficient forward solver for the implicit half (1, 2) and one AD pass for the explicit half (3). Gradient extraction… view at source ↗
Figure 2
Figure 2. Figure 2: Gradient-computation scaling. (a) Adjoint time vs. N. (b) Peak memory: UNROLL saturates the 128 GB budget. IFT-HYBRID is the only viable route to large-scale DARSM training. Empirical scaling of the hybrid adjoint. Section 3.3 argued that unrolled backpropagation cannot scale on stiff systems; we verify this empirically. We compare three gradient calculation routes for grids N ∈ {162 , 322 , 642 , 1282}: (… view at source ↗
Figure 3
Figure 3. Figure 3: Streamwise velocity W/Ub (filled contours) and in-plane secondary flow (white streamlines) at the test case Reb=11386. DARSM recovers the symmetric eight-vortex secondary-flow topology of the DNS (ground truth). TBNN, PINN, DeepONet, FIML, and the additive source-term closure produce distorted streamlines or weaker secondary motion. component-normalised velocity error J (θ) = 1 2 X 3 i=1 wi |Ω| Z Ω [PITH_… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution shift in the (S, Ω)-invariant feature space the a-priori network consumes. All invariants normalised per case (τ=1/ωRANS). Top: joint distributions at Reb=5000 of the strain pair (IIS,IIIS) and the rotation/coupling pair (IIΩ,IV), at the DNS state (training) and the converged default-RANS state (deployment). Bottom: 1-D marginals of all five invariants pooled across Reb ∈ {5000, 7000, 11386, 4… view at source ↗
Figure 5
Figure 5. Figure 5: Pointwise relative Frobenius error of the a-priori TBNN-Ling anisotropy [Ling et al., 2016] [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cross-stream velocity [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Periodic-hills cross-sections at Reb=5600 for five hill shapes α ∈ {0.5, 0.75, 1.0, 1.25, 1.5} (columns); the training case is α=1.0, the rest are test. Rows 1–3: streamwise velocity U/Ub (DNS, default EARSM, DARSM). Rows 4–6: wall-normal velocity V /Ub (same sources). Solid black contours mark each model’s U, V =0 separation lines; the dashed white contour re-draws the DNS reference on the model rows. unl… view at source ↗
Figure 8
Figure 8. Figure 8: Square-duct forward-solver validation. Top: streamwise velocity W/Ub (colour) with in-plane (u, v) secondary-flow streamlines, showing the characteristic eight-vortex pattern in the DNS [Vinuesa et al., 2018] at Reb = 12000 (left) and in our WJ-EARSM solution (right). Bottom: profiles of the streamwise W and cross-plane V components along the corner diagonal, compared to BSL-EARSM reference data [Menter, F… view at source ↗
Figure 9
Figure 9. Figure 9: Periodic-hills forward-solver validation at [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: End-to-end finite-difference verification of the hybrid adjoint gradient. Pointwise relative [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Velocity-only loss Jvel over outer iterations for one fold. Left: square duct (Reynolds split, H=10, BFGS). Right: periodic hills (shape-generalisation split, H=10, BFGS). Vertical dashed line marks the best-val epoch. Train decreases monotonically; val minimises then drifts. 0 25 50 75 100 125 150 outer iteration 4 × 10 −3 5 × 10 −3 6 × 10 −3 7 × 10 −3  val vel BFGS Adam [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 12
Figure 12. Figure 12: Validation loss trajectory for BFGS (used throughout) vs. Adam, matched architecture [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Learned closure corrections at Reb=5000: NN output for each of the seven k–ω/EARSM coefficients, expressed as fractional deviation from the default value. Red = pushed above default, blue = pushed below, white = default. The closure is spatially varying and concentrates its corrections near walls and along corner bisectors, exactly where secondary flow is generated. β * β0 γ σk σω c1 c2 −0.02 0.00 0.02 −0… view at source ↗
Figure 14
Figure 14. Figure 14: Learned closure corrections on periodic hills at [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
read the original abstract

Turbulence is ubiquitous in engineering and science, yet direct simulation is prohibitively expensive. The Reynolds-averaged Navier-Stokes (RANS) equations provide savings exceeding ten orders of magnitude but introduce unclosed terms (the closure problem). Offline-trained machine-learning (ML) closures suffer distribution shift in predictive simulations, while ML methods that bypass the governing equations struggle to generalise from scarce high-fidelity data. We develop a physics-derived deep learning closure model for RANS, the Deep Algebraic Reynolds Stress Model (DARSM), which can be trained on small datasets and accurately generalise across Reynolds numbers, to unseen geometries, and to different flow regimes. A neural network maps flow invariants to empirical parameters in an implicit algebraic Reynolds stress equation, derived from the Reynolds stress transport equations under the weak-equilibrium assumption, imposing physics-based structure on the ML closure. End-to-end optimisation through the governing PDEs and the coupled implicit closure eliminates distribution shift, but both unrolled and implicit automatic differentiation fail on the stiff coupled solver. We derive adjoint equations that exploit the solver's implicit-explicit structure for efficient optimisation. On canonical square-duct and periodic-hill benchmarks, DARSM reduces average test velocity error over baseline RANS by $2$-$4\times$ across Reynolds number, geometries, and flow regimes, with peak case-level reductions of $12\times$. The model trained on attached, anisotropy-dominated flows (square duct) accurately generalises without retraining to separated flows (periodic hills), a regime change in the underlying physics. DARSM also outperforms five established ML methods: offline training, tensor-basis neural networks, field-inversion machine learning, DeepONets, and physics-informed neural networks.

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.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the Deep Algebraic Reynolds Stress Model (DARSM), which embeds a neural network inside an implicit algebraic Reynolds stress closure derived from the Reynolds stress transport equations under the weak-equilibrium assumption. A NN maps invariants to closure coefficients; the model is trained end-to-end by minimizing the RANS residual on square-duct data and is reported to generalize without retraining to periodic-hill flows, yielding 2–4× average velocity-error reductions (peak 12×) relative to baseline RANS while outperforming five other ML closures.

Significance. If the cross-regime generalization holds, the work would be significant for turbulence modeling: it supplies a physics-structured algebraic form that mitigates distribution shift, demonstrates training on small attached-flow datasets, and introduces an adjoint formulation that exploits the implicit-explicit solver structure for efficient optimization. These elements address long-standing obstacles in ML-RANS closures.

major comments (3)
  1. [abstract and benchmark results] The headline generalization claim (abstract) that a model trained only on attached square-duct flows accurately predicts separated periodic-hill flows rests on the weak-equilibrium assumption remaining valid in the separation bubble. No direct comparison of the resulting algebraic Reynolds stresses against either full Reynolds-stress transport solutions or DNS data is supplied for the hill geometry, leaving open the possibility that observed velocity improvements arise from other factors rather than the structured ML closure.
  2. [abstract and evaluation sections] Quantitative performance statements (2–4× average error reduction, 12× peak) are given without reported error bars, explicit train/test splits, baseline implementation details, or sensitivity tests to the weak-equilibrium assumption (abstract). These omissions make it impossible to judge whether the reported gains are statistically robust or sensitive to the regime change.
  3. [model derivation] The derivation of the implicit algebraic form (via weak equilibrium) is presented as the key physics constraint, yet the manuscript does not quantify the magnitude of the neglected convective and diffusive terms in the periodic-hill separation region; if those terms are O(1) relative to production and pressure-strain, the algebraic closure itself becomes inconsistent with the underlying transport equations in the test regime.
minor comments (2)
  1. [model formulation] Notation for the invariants and the neural-network output coefficients should be unified between the derivation and the results tables to avoid reader confusion.
  2. [figures] Figure captions for the velocity and stress profiles should explicitly state the Reynolds numbers and the precise definition of the reported error norms.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments raise important points on validation of the weak-equilibrium assumption, statistical robustness, and consistency of the algebraic closure. We respond point-by-point below and will make revisions to address the concerns.

read point-by-point responses

  1. Referee: [abstract and benchmark results] The headline generalization claim (abstract) that a model trained only on attached square-duct flows accurately predicts separated periodic-hill flows rests on the weak-equilibrium assumption remaining valid in the separation bubble. No direct comparison of the resulting algebraic Reynolds stresses against either full Reynolds-stress transport solutions or DNS data is supplied for the hill geometry, leaving open the possibility that observed velocity improvements arise from other factors rather than the structured ML closure.

    Authors: We agree that a direct comparison of the modeled algebraic Reynolds stresses to DNS data on the periodic hill would strengthen the evidence that improvements stem from the structured closure rather than other factors. While mean-velocity accuracy is the primary RANS objective and our cross-method comparisons support the claim, we will add a new subsection and figure comparing predicted Reynolds-stress components against available DNS data in the separation region. revision: yes

  2. Referee: [abstract and evaluation sections] Quantitative performance statements (2–4× average error reduction, 12× peak) are given without reported error bars, explicit train/test splits, baseline implementation details, or sensitivity tests to the weak-equilibrium assumption (abstract). These omissions make it impossible to judge whether the reported gains are statistically robust or sensitive to the regime change.

    Authors: The reported benchmarks are deterministic fixed-mesh simulations, so ensemble-based error bars do not apply. We will expand the evaluation section to explicitly document the train/test splits (square-duct cases for training, periodic-hill cases for testing), provide full baseline implementation details and references, and include a sensitivity analysis to the weak-equilibrium assumption by comparing results with and without selected neglected terms. revision: yes

  3. Referee: [model derivation] The derivation of the implicit algebraic form (via weak equilibrium) is presented as the key physics constraint, yet the manuscript does not quantify the magnitude of the neglected convective and diffusive terms in the periodic-hill separation region; if those terms are O(1) relative to production and pressure-strain, the algebraic closure itself becomes inconsistent with the underlying transport equations in the test regime.

    Authors: We concur that quantifying the neglected terms is necessary to evaluate consistency of the algebraic form in the separated regime. Using the DNS data already employed for the periodic-hill benchmark, we will compute and report the relative magnitudes of convective, diffusive, and production/pressure-strain terms inside the separation bubble and discuss implications for the weak-equilibrium assumption. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses external modeling assumption and cross-regime test

full rationale

The algebraic form is obtained by applying the weak-equilibrium assumption to the Reynolds stress transport equations, a standard external modeling step not derived from the present data or NN. The NN maps invariants to coefficients and is trained on square-duct flows then evaluated on periodic-hill flows (different geometry and regime), so the reported generalization is not forced by construction. No self-citation chain, fitted-input-as-prediction, or renaming of known results appears in the provided text. The end-to-end PDE optimization is the intended training procedure rather than a tautological reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the weak-equilibrium assumption used to obtain the algebraic closure and on the empirical mapping learned by the neural network; no new physical entities are postulated.

free parameters (1)
  • Neural network weights and biases
    The parameters of the network that map flow invariants to the empirical coefficients of the algebraic Reynolds stress equation are determined by optimization against reference data.
axioms (1)
  • domain assumption Weak-equilibrium assumption
    Invoked to reduce the Reynolds stress transport equations to an implicit algebraic form that the neural network then parameterizes.

pith-pipeline@v0.9.1-grok · 5847 in / 1470 out tokens · 47736 ms · 2026-06-29T20:05:53.760307+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 10 canonical work pages

  1. [1]

    A probabilistic, data-driven closure model for RANS simulations with aleatoric, model uncertainty

    Atul Agrawal and Phaedon-Stelios Koutsourelakis. A probabilistic, data-driven closure model for RANS simulations with aleatoric, model uncertainty. Journal of Computational Physics, 508: 0 113024, 2024

    2024

  2. [2]

    Albring, Max Sagebaum, and Nicolas R

    Tim A. Albring, Max Sagebaum, and Nicolas R. Gauger. Development of a consistent discrete adjoint solver in an evolving aerodynamic design framework. In 16th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, page 3240, 2015

    2015

  3. [3]

    Breuer, N

    M. Breuer, N. Peller, Ch. Rapp, and M. Manhart. Flow over periodic hills -- numerical and experimental study in a wide range of R eynolds numbers. Computers & Fluids, 38 0 (2): 0 433--457, 2009

    2009

  4. [4]

    T. J. Craft, B. E. Launder, and K. Suga. Development and application of a cubic eddy-viscosity model of turbulence. International Journal of Heat and Fluid Flow, 17 0 (2): 0 108--115, 1996. doi:10.1016/0142-727X(95)00079-6

    work page doi:10.1016/0142-727x(95)00079-6 1996

  5. [5]

    MacArt, and Justin Sirignano

    Daniel Dehtyriov, Jonathan F. MacArt, and Justin Sirignano. oRANS: Online optimisation of RANS machine learning model with embedded DNS data generation . arXiv preprint arXiv:2510.02982, 2024

    arXiv 2024

  6. [6]

    Turbulence modeling in the age of data

    Karthik Duraisamy, Gianluca Iaccarino, and Heng Xiao. Turbulence modeling in the age of data. Annual Review of Fluid Mechanics, 51: 0 357--377, 2019

    2019

  7. [7]

    Yuan Fang, Yaomin Zhao, Fabian Waschkowski, Andrew S. H. Ooi, and Richard D. Sandberg. Toward more general turbulence models via multicase computational-fluid-dynamics-driven training. AIAA Journal, 61 0 (5): 0 2100--2115, 2023

    2023

  8. [8]

    Ferziger and Milovan Peri \'c

    Joel H. Ferziger and Milovan Peri \'c . Computational Methods for Fluid Dynamics. Springer, 3rd edition, 2002

    2002

  9. [9]

    A posteriori learning for quasi-geostrophic turbulence parametrization

    Hugo Frezat, Julien Le Sommer, Ronan Fablet, Guillaume Balarac, and Redouane Lguensat. A posteriori learning for quasi-geostrophic turbulence parametrization. Journal of Advances in Modeling Earth Systems, 14 0 (11): 0 e2022MS003124, 2022

    2022

  10. [10]

    An introduction to the adjoint approach to design

    Michael B Giles and Niles A Pierce. An introduction to the adjoint approach to design. Flow, Turbulence and Combustion, 65 0 (3): 0 393--415, 2000

    2000

  11. [11]

    Learning to control PDE s with differentiable physics

    Philipp Holl, Vladlen Koltun, and Nils Thuerey. Learning to control PDE s with differentiable physics. In International Conference on Learning Representations, 2020

    2020

  12. [12]

    Holland, James D

    Jonathan R. Holland, James D. Baeder, and Karthik Duraisamy. Towards integrated field inversion and machine learning with embedded neural networks for RANS modeling. In AIAA Scitech 2019 Forum, page 1884, 2019

    2019

  13. [13]

    Aerodynamic design via control theory

    Antony Jameson. Aerodynamic design via control theory. Journal of Scientific Computing, 3 0 (3): 0 233--260, 1988

    1988

  14. [14]

    Stiff-PINN : Physics-informed neural network for stiff chemical kinetics

    Weiqi Ji, Weilun Qiu, Zhiyu Shi, Shaowu Pan, and Sili Deng. Stiff-PINN : Physics-informed neural network for stiff chemical kinetics. The Journal of Physical Chemistry A, 125 0 (36): 0 8098--8106, 2021

    2021

  15. [15]

    Neural network-augmented eddy viscosity closures for turbulent premixed jet flames

    Priyesh Kakka and Jonathan F MacArt. Neural network-augmented eddy viscosity closures for turbulent premixed jet flames. Combustion and Flame, 278: 0 114241, 2025

    2025

  16. [16]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2015

    2015

  17. [17]

    Machine learning--accelerated computational fluid dynamics

    Dmitrii Kochkov, Jamie A Smith, Ayya Alieva, Qing Wang, Michael P Brenner, and Stephan Hoyer. Machine learning--accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118 0 (21), 2021

    2021

  18. [18]

    Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W

    Aditi S. Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W. Mahoney. Characterizing possible failure modes in physics-informed neural networks. In Advances in Neural Information Processing Systems, 2021

    2021

  19. [19]

    B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a R eynolds-stress turbulence closure. Journal of Fluid Mechanics, 68 0 (3): 0 537--566, 1975. doi:10.1017/S0022112075001814

    work page doi:10.1017/s0022112075001814 1975

  20. [20]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhatt, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021

    2021

  21. [21]

    Reynolds averaged turbulence modelling using deep neural networks with embedded invariance

    Julia Ling, Andrew Kurzawski, and Jeremy Templeton. Reynolds averaged turbulence modelling using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807: 0 155--166, 2016

    2016

  22. [22]

    Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons

    Björn List, Li-Wei Chen, and Nils Thuerey. Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons. Journal of Fluid Mechanics, 949: 0 A25, 2022. doi:10.1017/jfm.2022.738

    work page doi:10.1017/jfm.2022.738 2022

  23. [23]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3 0 (3): 0 218--229, 2021

    2021

  24. [24]

    Warp: A python framework for high-performance GPU simulation and graphics, 2022

    Miles Macklin. Warp: A python framework for high-performance GPU simulation and graphics, 2022. URL https://github.com/NVIDIA/warp. NVIDIA

    2022

  25. [25]

    Menter, F. R. , Garbaruk, A. V. , and Egorov, Y. Explicit algebraic reynolds stress models for anisotropic wall-bounded flows, 2012. URL https://doi.org/10.1051/eucass/201203089

    work page doi:10.1051/eucass/201203089 2012

  26. [26]

    Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, 2nd edition, 2006

    2006

  27. [27]

    Automating turbulence modelling by multi-agent reinforcement learning

    Guido Novati, Hugues Lascombes de Laroussilhe, and Petros Koumoutsakos. Automating turbulence modelling by multi-agent reinforcement learning. Nature Machine Intelligence, 3 0 (1): 0 87--96, 2021. doi:10.1038/s42256-020-00272-0

    work page doi:10.1038/s42256-020-00272-0 2021

  28. [28]

    A paradigm for data-driven predictive modeling using field inversion and machine learning

    Eric J Parish and Karthik Duraisamy. A paradigm for data-driven predictive modeling using field inversion and machine learning. Journal of Computational Physics, 305: 0 758--774, 2016

    2016

  29. [29]

    Turbulence and secondary motions in square duct flow

    Sergio Pirozzoli, Davide Modesti, Paolo Orlandi, and Francesco Grasso. Turbulence and secondary motions in square duct flow. Journal of Fluid Mechanics, 840: 0 631--655, 2018

    2018

  30. [30]

    Stephen B. Pope. A more general effective-viscosity hypothesis. Journal of Fluid Mechanics, 72 0 (2): 0 331--340, 1975

    1975

  31. [31]

    Stephen B. Pope. Turbulent Flows. Cambridge University Press, 2000. doi:10.1017/CBO9780511840531

    work page doi:10.1017/cbo9780511840531 2000

  32. [32]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378: 0 686--707, 2019

    2019

  33. [33]

    Global convergence of adjoint-optimized neural pdes

    Konstantin Riedl, Justin Sirignano, and Konstantinos Spiliopoulos. Global convergence of adjoint-optimized neural pdes. Journal of Machine Learning Research, 26 0 (295): 0 1--94, 2025. URL http://jmlr.org/papers/v26/25-1383.html

    2025

  34. [34]

    W. Rodi. A new algebraic relation for calculating the Reynolds stresses. ZAMM - Zeitschrift f \"u r Angewandte Mathematik und Mechanik , 56 0 (S1): 0 T219--T221, 1976

    1976

  35. [35]

    J. C. Rotta. Statistische theorie nichthomogener turbulenz. Zeitschrift f \"u r Physik , 129 0 (6): 0 547--572, 1951. doi:10.1007/BF01330059

    work page doi:10.1007/bf01330059 1951

  36. [36]

    Iterative Methods for Sparse Linear Systems

    Yousef Saad. Iterative Methods for Sparse Linear Systems. SIAM, 2nd edition, 2003

    2003

  37. [37]

    Benjamin Sanderse, Panos Stinis, Romit Maulik, and Shady E. Ahmed. Scientific machine learning for closure models in multiscale problems: A review. Foundations of Data Science, 7 0 (1): 0 298--337, 2025. doi:10.3934/fods.2024043

    work page doi:10.3934/fods.2024043 2025

  38. [38]

    Dwight, and Paola Cinnella

    Martin Schmelzer, Richard P. Dwight, and Paola Cinnella. Discovery of algebraic Reynolds-Stress models using sparse symbolic regression. Flow, Turbulence and Combustion, 104 0 (2): 0 579--603, 2020. doi:10.1007/s10494-019-00089-x

    work page doi:10.1007/s10494-019-00089-x 2020

  39. [39]

    Differentiable turbulence: Closure as a partial differential equation constrained optimization

    Varun Shankar, Dibyajyoti Chakraborty, Venkatasubramanian Viswanathan, and Romit Maulik. Differentiable turbulence: Closure as a partial differential equation constrained optimization. arXiv preprint arXiv:2307.03683, 2023

    arXiv 2023

  40. [40]

    DPM : A deep learning pde augmentation method with application to large-eddy simulation

    Justin Sirignano, Jonathan F MacArt, and Jonathan B Freund. DPM : A deep learning pde augmentation method with application to large-eddy simulation. Journal of Computational Physics, 423: 0 109811, 2020

    2020

  41. [41]

    MacArt, and Konstantinos Spiliopoulos

    Justin Sirignano, Jonathan F. MacArt, and Konstantinos Spiliopoulos. PDE -constrained models with neural network terms: Optimization and global convergence. Journal of Computational Physics, 481: 0 112016, 2023

    2023

  42. [42]

    Michel \'e n Str \"o fer and Heng Xiao

    Carlos A. Michel \'e n Str \"o fer and Heng Xiao. End-to-end differentiable learning of turbulence models from indirect observations. Theoretical and Applied Mechanics Letters, 11 0 (4): 0 100280, 2021

    2021

  43. [43]

    Solver-in-the-loop: Learning from differentiable physics to interact with iterative PDE -solvers

    Kiwon Um, Robert Brand, Yun Raymond Fei, Philipp Holl, and Nils Thuerey. Solver-in-the-loop: Learning from differentiable physics to interact with iterative PDE -solvers. In Advances in Neural Information Processing Systems, volume 33, 2020

    2020

  44. [44]

    Vinuesa, P

    R. Vinuesa, P. Schlatter, and H. M. Nagib. Secondary flow in turbulent ducts with increasing aspect ratio. Phys. Rev. Fluids, 3: 0 054606, 2018

    2018

  45. [45]

    An explicit algebraic Reynolds stress model for incompressible and compressible turbulent flows

    Stefan Wallin and Arne V Johansson. An explicit algebraic Reynolds stress model for incompressible and compressible turbulent flows. Journal of Fluid Mechanics, 403: 0 89--132, 2000

    2000

  46. [46]

    When and why PINNs fail to train: A neural tangent kernel perspective

    Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449: 0 110768, 2022

    2022

  47. [47]

    Weatheritt and R.D

    J. Weatheritt and R.D. Sandberg. The development of algebraic stress models using a novel evolutionary algorithm. International Journal of Heat and Fluid Flow, 68: 0 298--318, 2017

    2017

  48. [48]

    Reassessment of the scale-determining equation for advanced turbulence models

    David C Wilcox. Reassessment of the scale-determining equation for advanced turbulence models. AIAA Journal, 26 0 (11): 0 1299--1310, 1988

    1988

  49. [49]

    Reynolds-averaged navier--stokes equations with explicit data-driven reynolds stress closure can be ill-conditioned

    Jinlong Wu, Heng Xiao, Rui Sun, and Qiqi Wang. Reynolds-averaged navier--stokes equations with explicit data-driven reynolds stress closure can be ill-conditioned. Journal of Fluid Mechanics, 869: 0 553--586, 2019. doi:10.1017/jfm.2019.205

    work page doi:10.1017/jfm.2019.205 2019

  50. [50]

    Flows over periodic hills of parameterized geometries: A dataset for data-driven turbulence modeling from direct simulations

    Heng Xiao, Jin-Long Wu, Sylvain Laizet, and Lian Duan. Flows over periodic hills of parameterized geometries: A dataset for data-driven turbulence modeling from direct simulations. Computers & Fluids, 200: 0 104431, 2020

    2020