Abstract
In this paper, we propose normalizing flows (NF) as a novel probability density function (PDF) turbulence model (NF-PDF model) for the Reynolds-averaged Navier-Stokes (RANS) equations. We propose to use normalizing flows in two different ways: firstly, as a direct model for the Reynolds stress tensor, and secondly as a second-moment closure model, i.e. for modeling the pressure-strain and dissipation tensor in the Reynolds stress transport equation. In classical PDF closure models, a stochastic differential equation has to be modeled and solved to obtain samples of turbulent quantities. The NF-PDF closure model allows for direct sampling from the underlying probability density functions of fluctuating turbulent quantities, such that ensemble-averaged quantities can then be computed. To illustrate this approach we demonstrate an application for the canonical case of homogeneous shear turbulence. Training data are extracted from high-fidelity direct numerical simulations (DNS).
Introduction: The Challenge of Turbulence Modeling
Simulating turbulent flow remains one of the most significant challenges in computational physics. While Direct Numerical Simulation (DNS) offers high accuracy by resolving all scales of motion, its computational cost is prohibitive for industrial-scale applications. Consequently, engineers rely on Reynolds-averaged Navier-Stokes (RANS) equations, which model the mean flow by averaging out turbulent fluctuations.
However, this averaging process introduces the "closure problem," where new unknown terms, such as the Reynolds stress tensor, must be modeled. Traditional closures often assume isotropic turbulence, which fails in complex, real-world flows. Probability Density Function (PDF) methods provide a higher-fidelity alternative by tracking the distribution of fluctuations, but they typically require solving complex stochastic differential equations (SDEs) that are difficult to scale. This research explores Normalizing Flows (NFs) as a generative machine learning solution to simplify and accelerate this process.
Methodology: Normalizing Flows for PDF Closure
The proposed NF-PDF model utilizes Normalizing Flows—a class of generative models that transform a simple base distribution (e.g., Gaussian) into a complex target distribution through a sequence of invertible transformations. Specifically, the authors employ Real-valued Non-Volume Preserving (Real NVP) architectures.
Two modeling strategies are investigated:
- Direct Reynolds Stress Modeling: The NF learns the PDF of fluctuating velocities ($u'_i$), enabling rapid sampling to compute the Reynolds stress tensor.
- Second-Moment Closure: The NF models unclosed terms within the Reynolds stress transport equation, specifically the pressure-strain correlation and the dissipation tensor.
To account for the temporal evolution of turbulence, Conditional Normalizing Flows are used, where the model is conditioned on time ($St$), allowing it to predict the trajectory of the PDF as the flow develops.
Experimental Setup and Data
The models were trained using data from 3D DNS of Homogeneous Shear Turbulence (HST). HST serves as a canonical benchmark, capturing the essential physics of shear—where fluid layers slide at different speeds—while remaining computationally manageable. The DNS data, generated via a parallelized pseudo-spectral code, provides the ground truth for fluctuating velocities, pressures, and velocity gradients.
Key Insights: Beyond Gaussian Assumptions
A critical advantage of the NF approach is its ability to capture the non-Gaussian nature of turbulence. Real-world turbulent fluctuations often exhibit "heavy tails" that standard models miss.
As shown in Figure 3, the NF model accurately reproduces the PDF of velocity fluctuations observed in DNS, significantly outperforming a simple Gaussian approximation. By learning the full joint distribution, the NF naturally accounts for anisotropy (direction-dependent behavior), a common failure point for traditional RANS closures. Furthermore, by bypassing the need to solve SDEs, NFs allow for direct sampling, which is significantly more efficient for computing ensemble averages.
Results and Performance
The NF-PDF model demonstrates high precision in predicting turbulent quantities. For instance, the velocity anisotropy tensor component $b_{11}$ was predicted with remarkable accuracy (0.245 in both DNS and NF model).
Figure 5 illustrates the model's performance in second-moment closure. The NF accurately captures the dissipation term ($\epsilon_{ij}$) across various components. While the pressure-strain correlation ($\Pi_{ij}$)—a notoriously difficult term representing non-local interactions—showed minor deviations at early time steps, the overall agreement with DNS data is strong, particularly as the flow matures.
Conclusion and Future Directions
This work establishes Normalizing Flows as a powerful tool for data-driven turbulence modeling. By enabling direct sampling from learned PDFs, the NF-PDF model offers a path toward high-fidelity RANS closures without the computational burden of traditional PDF methods.
Future research will focus on extending this framework to inhomogeneous flows, where turbulence varies spatially (e.g., near solid boundaries). Additionally, improving the fidelity of pressure-strain predictions and investigating the model's generalizability across different Reynolds numbers remain key priorities for developing robust, industry-ready turbulence models.
Summary updated