A data-driven approach for 2D vorticity PDF equations by a new conditional average estimation

This paper presents a hybrid data-driven method that utilizes DNS data and a sampling estimator to numerically solve dimensionally reduced, linear kinetic transport equations for one-point and two-point vorticity probability density functions in two-dimensional homogeneous isotropic turbulence.

The Gist

Imagine you are trying to predict the weather, but instead of clouds and rain, you are tracking the swirling eddies of a fluid, like cream mixing into coffee or smoke rising from a candle. In physics, these swirls are called vortices, and the measure of how fast they spin is called vorticity.

The paper you provided is about a new, clever way to predict how these swirls behave over time, specifically in a flat, 2D world (like a thin layer of water).

Here is the breakdown of their work using simple analogies:

1. The Problem: The "Too Many Variables" Puzzle

In the real world, fluids are messy. They have billions of tiny particles interacting. Trying to write a single equation that predicts the exact path of every single swirl is impossible. It's like trying to predict the exact path of every single grain of sand in a beach storm.

Scientists usually use "Probability Density Functions" (PDFs). Think of a PDF not as a map of one specific swirl, but as a weather forecast for the whole crowd. Instead of saying "Swirl A is here," it says, "There is a 10% chance a swirl is spinning this fast, and a 5% chance it's spinning that fast."

The problem is that the equations for these "crowd forecasts" are incomplete. They have a missing piece, a "black box" term that depends on the complex interactions of the fluid. It's like having a recipe for a cake, but the instructions say, "Add the secret ingredient," without telling you what it is or how much to add.

2. The Solution: A Hybrid Detective

The authors (Huang, Görtz, et al.) came up with a "hybrid" method. They didn't try to guess the secret ingredient mathematically. Instead, they used a data-driven detective approach.

• The Simulation (The Crime Scene): They ran massive, high-speed computer simulations (called DNS) of the fluid. This generated a huge database of "snapshots" showing exactly how the fluid behaved.
• The Estimator (The Detective): They needed to figure out that missing "black box" term. They used a statistical tool called the Nadaraya-Watson estimator.
  • The Analogy: Imagine you want to know how fast a car usually drives when it's raining. You don't need to know the physics of the engine. You just look at your database of past trips. You find all the times it was raining, look at the speed of the cars in those specific moments, and take the average.
  • In this paper, they looked at specific moments when the fluid had a certain spin speed, found all the similar moments in their simulation data, and calculated the average behavior of the "missing term" for those moments.

3. The Process: The "Crowd Control" Method

Once they figured out the missing term using their data, they plugged it back into the equation.

• The Equation: The equation describes how the "crowd forecast" (the PDF) moves and changes over time. It's like a conveyor belt moving the probability of different spin speeds.
• The Method: They solved this equation by tracking "characteristics."
  • The Analogy: Imagine a river (the equation) carrying leaves (the probability). Instead of trying to calculate the water's movement everywhere at once, they just tracked where specific leaves started and where they ended up. This made the math much faster and more stable.

4. The Results: Two Different Worlds

They tested their method on two scenarios:

A. The Fading Party (Decaying Turbulence)

• Scenario: You stir the coffee and then stop. The swirls slowly die out due to friction (viscosity).
• Result: The method perfectly predicted how the swirls changed. At first, the spins were random (Gaussian), but as time went on, the "crowd" of swirls became more concentrated around zero (calm), with a few extreme outliers. The method captured this "fading" perfectly.

B. The Never-Ending Party (Forced Turbulence)

• Scenario: You keep stirring the coffee at a steady pace to keep the swirls going forever.
• Result: The method predicted a "steady state" where the swirls never die out. It correctly showed that the energy added by stirring balanced the energy lost to friction. It even revealed how the stirring and friction fought each other to keep the system stable.

Why This Matters

This paper is a bridge between pure theory and big data.

• Old Way: Try to guess the missing math terms using complex theories (often fails).
• New Way: Use the computer to "observe" the fluid, learn the missing rules from the data, and then use those rules to predict the future.

In a nutshell: They built a machine that learns the "rules of the game" by watching a million simulations, and then uses those learned rules to accurately predict how fluid turbulence evolves, without needing to solve the impossible math of every single particle. It's like teaching a computer to play chess by watching grandmasters, rather than trying to program the laws of physics for every piece.

Technical Summary

1. Problem Statement

The statistical description of turbulence remains a fundamental challenge due to the nonlinear, multiscale nature of the Navier-Stokes equations. The Lundgren-Monin-Novikov (LMN) hierarchy provides a formally exact framework for describing flow variables via multi-point probability density functions (PDFs). However, this hierarchy is unclosed; the evolution equations for lower-order PDFs depend on conditional averages of higher-order dynamical quantities (e.g., pressure gradients, dissipation, forcing).

Specifically for Two-Dimensional Homogeneous Isotropic Turbulence (2D HIT), the dynamics are governed by a scalar vorticity equation. While the absence of vortex stretching simplifies the physics compared to 3D turbulence, the statistical structure is still complex due to the dual energy/enstrophy cascade and the formation of coherent vortices. The core problem addressed in this paper is the predictive numerical closure of the vorticity PDF hierarchy. Traditional modeling often relies on Gaussian approximations or heuristic closures, which fail to capture non-Gaussian statistics and intermittency. The authors aim to solve the PDF transport equations directly by estimating the unclosed conditional average terms using data from Direct Numerical Simulations (DNS).

2. Methodology

The authors propose a hybrid data-driven method that combines the rigorous derivation of PDF transport equations with non-parametric statistical estimation from DNS data.

A. Theoretical Framework

• Derivation: Starting from the 2D vorticity transport equation, the authors derive dimensionally reduced governing equations for one-point and two-point PDFs under assumptions of homogeneity and isotropy.
• Equation Form: The resulting equations are linear kinetic transport equations in the sample space (vorticity space).
  • For decaying turbulence: $\frac{\partial f_1}{\partial t} = -\frac{1}{Re} \frac{\partial}{\partial \Omega_1} (\langle \Delta \omega | \Omega_1 \rangle f_1)$
  • For forced turbulence: Additional terms for forcing and large-scale damping are included as conditional averages.
• Unclosed Terms: The drift terms in these equations are conditional expectations (e.g., $\langle \Delta \omega | \omega = \Omega_1 \rangle$), which represent the average viscous dissipation or forcing effect given a specific vorticity value.

B. Numerical Strategy

1. Data Generation (DNS):
   • Simulations are performed using a pseudo-spectral solver on a doubly periodic domain ($[0, 2\pi]^2$).
   • Cases include decaying HIT and forced HIT (using linear forcing at small scales and large-scale damping) at Reynolds numbers $Re=200$ and $Re=360$.
2. Conditional Average Estimation:
   • Instead of modeling the conditional averages, they are estimated directly from DNS snapshots.
   • Estimator: A Nadaraya-Watson kernel estimator is used. The authors specifically employ a simple indicator (box) kernel, which corresponds to a local binning procedure in vorticity space.
   • Sampling: DNS grid points at a fixed time are treated as independent samples. The method allows for systematic convergence studies by varying the number of samples used in the estimator.
3. Solving the Transport Equation:
   • The PDF transport equation is reformulated in terms of the Cumulative Distribution Function (CDF), $F_1(\Omega_1, t)$.
   • The CDF satisfies a first-order transport equation: $\frac{\partial F_1}{\partial t} + M(\Omega_1, t) \frac{\partial F_1}{\partial \Omega_1} = 0$, where $M$ is the drift coefficient derived from the estimated conditional averages.
   • The equation is solved using a characteristic-based method (method of characteristics), where the CDF is conserved along characteristic curves defined by the drift velocity.
   • The final PDF is recovered via finite differences of the CDF.

3. Key Contributions

• Hybrid Closure Strategy: The paper introduces a robust framework that avoids heuristic modeling of closure terms. Instead, it uses a "data-driven" approach where the unclosed conditional averages are approximated via kernel estimation from high-fidelity DNS data.
• Efficiency of Estimation: The use of the Nadaraya-Watson estimator with a box kernel is shown to be computationally efficient and robust. It enhances sample efficiency compared to constructing distributions directly from Monte Carlo samples.
• Extension to Two-Point Statistics: While the primary focus is on one-point PDFs, the theoretical derivation is generalized to two-point PDFs, offering a structure-based way to capture spatial correlations essential in turbulence.
• Validation of Kinetic Approach: The study demonstrates that the kinetic transport equation, when supplied with accurate conditional averages, can faithfully reproduce the complex statistical evolution of turbulence without solving the full Navier-Stokes equations.

4. Results

The method was tested on four DNS cases (decaying and forced at two Reynolds numbers).

• Decaying Turbulence:
  • The method accurately captured the transition from an initial near-Gaussian distribution to a distribution that is highly peaked at zero with heavy tails (indicating intermittency and coherent vortex formation).
  • Convergence: The $L_2$-error of the reconstructed CDF decreased monotonically as the number of samples in the estimator increased.
  • Accuracy: With sufficient samples ($>10^4$), the reconstructed PDF became nearly indistinguishable from the direct DNS statistics.
• Forced Turbulence:
  • The method successfully recovered the statistically stationary PDFs.
  • Drift Analysis: The conditional averages revealed the physical balance: the viscous term provided a dissipative drift towards zero, while the forcing term provided a compensating injection of fluctuations. The net drift was approximately zero in the stationary regime, explaining the time-invariant shape of the PDF.
  • The "forced" PDFs were significantly wider than decaying ones, reflecting the sustained injection of extreme vorticity events.
• Error Sources: The primary source of error was identified as the statistical noise in estimating the conditional averages, particularly at low sample counts or in the tails of the distribution. The error in the PDF ($f_1$) grew with time, while the error in the CDF ($F_1$) remained sublinear and bounded.

5. Significance

This work provides a natural bridge between data-driven closure strategies and the theoretical framework of turbulent flows.

• Predictive Capability: It demonstrates that PDF hierarchies can be solved predictively if the conditional averages are accurately estimated, bypassing the need for ad-hoc turbulence models.
• Physical Insight: By analyzing the conditional averages, the method offers direct insight into the interplay between coherent structures, viscous dissipation, and external forcing.
• Scalability: The approach is not limited to one-point statistics; the framework is explicitly designed to extend to multi-point PDFs, which is crucial for capturing spatial correlations and the full complexity of turbulence.
• Computational Efficiency: By solving a transport equation for the PDF rather than re-running full DNS, this method offers a potential pathway for reduced-order modeling of turbulence statistics, provided the conditional averages can be estimated efficiently.

In summary, the paper validates a hybrid numerical method that leverages DNS data to close the LMN hierarchy, successfully reproducing the statistical evolution of 2D turbulence and offering a new paradigm for turbulence modeling.