Turbulence Model Augmented PINNs

Updated 1 May 2026

Turbulence Model Augmented PINNs are neural surrogates that embed classical turbulence closures (e.g., RANS k–ε, k–ω) to simulate turbulent flows.
They fuse data, low-fidelity models, and continuous physical constraints via composite loss functions, enabling forward modeling and inverse inference.
These methods demonstrate significant error reductions and generalization across diverse geometries, with potential extensions to unsteady and complex flows.

Turbulence model augmented physics-informed neural networks (PINNs) constitute a class of data-driven methods designed to simulate, infer, or reconstruct turbulent flows by embedding established turbulence models (e.g., RANS k–ε, k–ω, Spalart–Allmaras) directly into the structure and loss functional of PINNs. Unlike generic PINNs, which may treat turbulence closure terms implicitly or as unknowns, these methods explicitly merge data, low-fidelity closures, and continuous physical constraints, enabling efficient, accurate, and generalizable surrogates for turbulent flow regimes in both direct simulation and inversion/data assimilation settings. The approach has found traction in forward surrogate modeling, inverse inference, parametric studies, and hybrid experimental–numerical data assimilation across diverse geometries and flow regimes, from canonical cylinders to parameterized engineering configurations.

1. Governing Equations and Embedded Closures

Turbulence model augmented PINNs target turbulent regimes where the Reynolds-averaged Navier–Stokes (RANS) equations or closely related mean-field equations serve as the backbone. The challenge is the closure of the Reynolds stress term, $\overline{u_i' u_j'}$ , which results from averaging the nonlinear advection term. Classical approaches introduce eddy-viscosity models, e.g., k–ε, k–ω, Spalart–Allmaras (SA), to close the system. The turbulence model introduces additional PDEs and algebraic closures into the PINN framework, tightly coupling the mean flow variables to turbulence quantities such as turbulent kinetic energy ( $k$ ), dissipation rate ( $\epsilon$ or $\omega$ ), or eddy viscosity ( $\nu_t$ ). The canonical formulation encompasses:

Continuity: $\nabla\cdot\mathbf{u} = 0$
Momentum: $\mathbf{u}\cdot\nabla\mathbf{u} = -\frac{1}{\rho}\nabla p + \nu\nabla^2\mathbf{u} - \nabla\cdot\boldsymbol{\tau}$
Turbulence closure (e.g., k–ε): PDEs for $k$ , $\epsilon$ , and eddy viscosity $\nu_t = C_\mu k^2/\epsilon$
SA model: additional PDE for $k$ 0 with $k$ 1

In advanced models, the closures themselves may be nonlinear (e.g., $k$ 2-type neural surrogates for Reynolds stress), or data-driven corrections to the standard coefficient structure (e.g., PINN–NN for variable turbulent diffusion coefficients) are learned and embedded (Ghosh et al., 2023, Jiang et al., 22 Mar 2025, Davidson, 16 Nov 2025, Zhang et al., 7 Oct 2025, Patel et al., 2023, Toma et al., 8 Jan 2026).

2. Network Architectures, Input Encodings, and Turbulence-Model Integration

A consistent structural motif across turbulence model augmented PINNs is the explicit parameterization of the full mean flow and turbulence model state as network outputs:

Fully connected MLPs are standard: depth 4–10 layers, 30–512 neurons/layer, $k$ 3 or ReLU activations.
Inputs: spatial coordinates (e.g., $k$ 4, $k$ 5), parametric variables (Reynolds number, geometry parameters, viscosity), sometimes geometry encodings (SDF, design parameters for airfoils).
Outputs: mean velocity $k$ 6, pressure $k$ 7, turbulence quantities $k$ 8 or $k$ 9, eddy viscosity $\epsilon$ 0, and in some cases explicit Reynolds stress divergence or additional closure-related variables.

In “geometry-aware” surrogates, the network ingests both local signed-distance and global shape parameters for generalization to unseen geometries (Ghosh et al., 2024).

In parametric and multi-condition surrogates (e.g., PT-PINNs), the input is extended to accommodate physical or design parameter spaces, enabling solution generalization across $\epsilon$ 1, expansion ratio, or analogous parameters (Jiang et al., 22 Mar 2025, Ghosh et al., 2023).

Neural surrogates may be designed to output either the turbulence model algebraic closure (e.g., $\epsilon$ 2) or treat closure coefficients as functions learned via auxiliary NNs, as in $\epsilon$ 3– $\epsilon$ 4–PINN–NNs (Davidson, 16 Nov 2025).

3. Loss Function Composition and Training Strategies

Turbulence model augmented PINNs utilize meticulously crafted composite loss functions, blending supervised data mismatch, boundary enforcement, and physics-informed PDE residuals. The total loss $\epsilon$ 5 typically follows:

$\epsilon$ 6

$\epsilon$ 7: MSE between network and ground-truth (CFD or experimental) values of velocity, pressure, turbulence quantities; often limited to a sparse set of domain points.
$\epsilon$ 8: MSE on boundary conditions; hard constraints may be imposed via augmented Lagrangian (especially for no-slip walls).
$\epsilon$ 9: MSE of residuals for continuity, momentum, and turbulence model equations, evaluated via automatic differentiation and collocation.
$\omega$ 0: Terms enforcing model-specific constraints, e.g., soft or hard enforcement of $\omega$ 1, or training of closure coefficients via NN predictions.
$\omega$ 2: Regularizations such as $\omega$ 3 weight penalties or solenoidal constraint penalizations for non-unique decomposition in RANS inversion.

Adaptive weighting is employed to balance disparate PDE components (e.g., physics residuals vs. data-fidelity), using strategies such as residual-based weights, increasing PDE-loss weights as data loss stagnates, and logarithmic loss on steep-gradient quantities (notably $\omega$ 4). Initial phases (pre-training) minimize $\omega$ 5 for rapid warm starts, before introducing $\omega$ 6 and joint optimization (Ghosh et al., 2023, Jiang et al., 22 Mar 2025, Ghosh et al., 2024).

Transfer learning and multi-stage schedules are used to further improve data efficiency and solution realism (RANS pretraining, followed by experimental data fine-tuning) (Toma et al., 8 Jan 2026).

4. Parametric, Inverse, and Generalization Capabilities

PINN-based surrogates augmented with turbulence models demonstrate broad generalization across parameter spaces:

Parameterization in Re, geometry, and inlet conditions: Training over a discrete set of conditions (e.g., multiple $\omega$ 7 or shape parameters) allows PINNs to interpolate to unseen cases, drastically reducing repeated CFD cost (Ghosh et al., 2023, Jiang et al., 22 Mar 2025, Ghosh et al., 2024).
Inverse inference/data assimilation: Turbulence model augmented PINNs recover hidden fields (pressure, eddy viscosity, Reynolds stresses) from sparse velocity data, leveraging the physics loss to reconstruct inaccessible states—surpassing classical RANS and mesh-based DA, especially in separation and high-gradient regions (Patel et al., 2023, Saldern et al., 2022).
Data-driven closure learning: PINNs can be used to infer universal or flow-specific turbulence closures (e.g., local closures for Reynolds-stress divergence in a cylinder wake), outperforming standard Boussinesq or analytic models, fitting even cross-regime “universal” Reynolds stress fields (Zhang et al., 7 Oct 2025).

5. Quantitative Validation and Performance Analysis

Multiple studies offer rigorous quantitative assessment:

Reference	Model/Closure	Nominal Error vs. CFD/Exp/DNS	Notes
(Ghosh et al., 2023)	RANS k–ε (parametric PINN)	Velocity RMSE: 0.014–0.153; p: 0.029–0.164	Near real-time inference; low data requirement
(Jiang et al., 22 Mar 2025)	RANS k–ω (PT-PINN)	Velocity error <6%; TKE peak error <10%	$\omega$ 8 surrogate speedup vs CFD; strong separation/wall fidelity
(Ghosh et al., 2024)	RANS k–ε (geo-parametric)	1–6% (velocity), 0.3–5% (pressure)	Generalizes to unseen NACA shapes and Reynolds numbers
(Patel et al., 2023, Toma et al., 8 Jan 2026)	SA	Up to 85% error reduction vs baseline RANS	PINN-DA-SA consistently outperforming variational DA
(Zhang et al., 7 Oct 2025)	Data-driven closure	Velocity <1% L2, Reynolds-stress 5–20%	Closure learned from sparse PIV, cross-regime generalization

Other studies (Davidson, 16 Nov 2025) show channel and boundary layer velocity profiles within 1% and skin-friction within 6% of DNS, and PT-PINNs outperform standard CFD for separated 3D configurations at a fraction of the computational cost.

6. Practical Limitations and Open Challenges

Current turbulence model augmented PINNs encounter several practical bottlenecks:

Time-averaged approximations: RANS-model based surrogates capture only mean flows; unsteady phenomena (vortex shedding, coherent structures) remain inaccessible unless time-dependent/LES closures are embedded (Ghosh et al., 2023).
Near-wall modeling: High-fidelity predictions near wall boundaries remain challenging, with errors concentrated in thin boundary layers. Proposed remedies include wall-function enforcement and mesh/point refinement in these regions (Ghosh et al., 2023, Jiang et al., 22 Mar 2025, Ghosh et al., 2024).
Training cost: Training wall-clock for high-dimensional, parametric, or 3D PINNs remains nontrivial (e.g., 39 hours for PT-PINNs on a 4090 GPU), though still orders of magnitude faster than exhaustive CFD sweeps (Jiang et al., 22 Mar 2025).
Model generalizability: When classical RANS closures misrepresent the true flow physics (e.g., strong anisotropy), the PINN may manifest high residuals or fail to reconcile data and physics (Saldern et al., 2022).
Balance of losses and optimization: Proper weighting and staged scheduling of data vs. physics losses are critical to avoid local minima and oscillatory or unphysical solutions, especially when coupling multiple tightly-coupled PDEs (Ghosh et al., 2023, Jiang et al., 22 Mar 2025).

7. Extensions and Future Directions

Active research is targeting several directions:

Augmenting closure complexity: Embedding more advanced nonlinear RANS models (e.g., Reynolds-stress transport, SST k–ω), or training closure coefficients as additional outputs or functions of local invariants (Davidson, 16 Nov 2025, Ghosh et al., 2023, Jiang et al., 22 Mar 2025).
Higher-dimensional and unsteady surrogates: Transition from 2D steady to 3D unsteady PINNs for complex dynamic separation and turbulent structures, with extensions to heat transfer and deeper parametric spaces (Jiang et al., 22 Mar 2025, Ghosh et al., 2024).
Hybrid and hierarchical models: Using PINNs for scale-separation (low/high- $\omega$ 9 networks), or as surrogates for subgrid-scale models (PINN-2 in (Kag et al., 2022)), and integrating with neural operator/XPINN/domain-decomposition frameworks to improve scalability.
Adaptive sampling: Refining collocation point distributions dynamically to focus model expressivity and residual minimization in critical regions (e.g., wakes, shear layers) (Ghosh et al., 2023).
Experimental hybridization and transfer learning: Systematic use of pretraining (e.g., on RANS/CFD fields) and transfer learning with sparse experimental data (PIV)—demonstrated to improve data efficiency and physical realism, especially when pressure and Reynolds stress data are absent (Toma et al., 8 Jan 2026).

Turbulence model augmented PINNs thus represent a convergent paradigm, merging physics-based modeling, data-driven learning, and NN surrogacy to address challenges in high-fidelity, generalizable, and physically consistent turbulent flow prediction and inference.