Multilevel radial basis function surrogates for noise-robust DSMC-CFD coupling

This paper introduces an enhanced "MMS-Sparse" hybrid framework that utilizes multilevel radial basis functions (RBFs) to provide noise-robust, automated, and geometrically flexible coupling between DSMC and CFD solvers for simulating rarefied gas flows.

The Gist

Imagine you are trying to predict how a crowd of people moves through a massive, complex shopping mall. You have two ways to do this, but both have major flaws.

Method 1: The "Micro" Approach (Counting every person).
You hire thousands of observers to track every single person’s exact step, speed, and direction. This is incredibly accurate, but it is exhausting and expensive. You’ll run out of money and time before you even finish mapping the food court. This is like DSMC (the particle-based method used in the paper).

Method 2: The "Macro" Approach (The "Flow" Theory).
Instead of tracking people, you assume the crowd moves like a smooth liquid. You use math to say, "If people enter here, they will generally flow there." This is fast and cheap, but it’s too simple. It fails to account for the "chaos" at the edges—like people bumping into walls, stopping at a kiosk, or swirling in a circle near a fountain. This is like CFD (the fluid dynamics method).

The Problem: The "Noisy" Middle Ground

Scientists want to combine these two: use the "Liquid Theory" for the big open hallways and the "Individual Tracking" for the crowded corners.

However, there is a huge problem: The "Noisy" Data. Because the "Individual Tracking" (DSMC) relies on random movements, the data it produces is "jittery" and "noisy"—like a shaky, handheld video. If you try to plug that shaky data directly into your smooth "Liquid Theory" math, the whole system crashes. It’s like trying to build a smooth marble floor using jagged, vibrating bricks.

The Solution: The "Smart Smoothing" Filter

This paper introduces a new way to bridge this gap using something called Multilevel Radial Basis Functions (RBFs) and Sparse Bayesian Learning.

Think of this new method as a High-Tech Digital Filter (like the "Portrait Mode" or "Noise Reduction" on your smartphone camera).

1. The Smart Filter (The RBFs): Instead of using one giant, clumsy brush to smooth out the data, the researchers use a "multilevel" set of brushes. They have huge, wide brushes to capture the general flow of the crowd, and tiny, fine-tipped brushes to capture the intricate swirls near the walls. This allows them to be flexible enough to handle any weirdly shaped room or complex corner.
2. The Intelligent Guessing (Bayesian Learning): The "Bayesian" part is like a very smart assistant. When the data is shaky and noisy, the assistant doesn't just blindly follow the jitters. Instead, it says, "I see a lot of shaking here, but based on what I know about how fluids move, I'm going to ignore the jitter and focus on the actual trend." It effectively "cleans" the data without losing the important physics.

Why does this matter?

By using this "Smart Filter," the researchers proved they could:

• Save Time: They don't have to track every "person" (particle) in the whole mall—only the ones in the tricky corners.
• Be Accurate: They get results that are almost as good as the expensive, slow method, but much faster.
• Handle Complexity: Because of those "multilevel brushes," this method can work in complicated shapes (like a "lid-driven cavity," which is basically a box where the top lid is sliding, creating complex swirls), not just simple straight pipes.

In short: They’ve created a way to take "shaky" microscopic data and turn it into "smooth" macroscopic maps, allowing scientists to simulate complex gas flows (like air moving around a spacecraft) much faster and more reliably than ever before.

Technical Summary

Technical Summary: Multilevel Radial Basis Function Surrogates for Noise-Robust DSMC-CFD Coupling

1. Problem Statement

Simulating rarefied gas flows (where the Knudsen number $Kn$ is significant) presents a computational dichotomy. The Direct Simulation Monte Carlo (DSMC) method is accurate across all regimes but is computationally prohibitive in the continuum and slip regimes due to high particle counts and small time steps required to manage statistical noise. Conversely, Computational Fluid Dynamics (CFD) based on Navier-Stokes equations is efficient but loses accuracy as the gas becomes more rarefied.

While hybrid DSMC-CFD coupling methods exist to combine these approaches, they face four critical challenges:

1. Noise Robustness: Statistical noise from DSMC can cause numerical instabilities in the CFD solver.
2. Physical Reliability: Denoising methods must ensure that transferred data satisfies conservation laws.
3. Automation: The coupling process should not require manual parameter tuning.
4. Geometric Flexibility: Existing surrogate-based hybrid methods often rely on global basis functions (like polynomials), limiting their use to simple, one-dimensional geometries.

2. Methodology

The authors propose an enhanced version of the Micro-Macro-Surrogate-Sparse (MMS-Sparse) framework. The methodology is structured as follows:

• Micro-Model (DSMC): Uses the SPARTA solver to generate particle-based data for macroscopic quantities (velocity $u$, pressure $p$, and stress tensor $\tau$).
• Surrogate Modeling (Sparse Bayesian Learning): Instead of simple filtering, the method uses Sequential Sparse Bayesian Learning (SSBL). This approach treats the DSMC data as a combination of a smooth signal and Gaussian noise. It automatically determines model complexity by pruning irrelevant basis functions, providing a noise-robust regression that avoids the overfitting common in frequentist (MLE) approaches.
• Novel Basis Functions (Multilevel RBFs): To solve the flexibility problem, the authors replace global polynomials with multilevel squared-exponential Radial Basis Functions (RBFs). These RBFs are organized into hierarchical levels with increasing resolution (from broad to fine details), allowing the model to capture complex, spatially two-dimensional flow features in arbitrary geometries.
• Macro-Model (CFD): The learned surrogates provide "smooth corrections" to the Navier-Stokes equations. These corrections are embedded into the OpenFOAM (icoFoam) solver as constitutive stress corrections ($\Phi$) and slip boundary conditions. To handle high $Kn$ regimes, a viscosity minimization technique is used to moderate the Newtonian contribution.

3. Key Contributions

• Spatial Generalization: Transitioned the MMS-Sparse framework from 1D channel flows to complex 2D flows (Lid-Driven Cavity).
• Multilevel RBF Integration: Introduced a hierarchical RBF structure that provides the local flexibility needed for complex geometries while maintaining the automation of the Bayesian framework.
• Robustness & Automation: Demonstrated a method that automatically handles noise reduction and model selection without manual intervention.
• Uncertainty Quantification: Utilized the Bayesian posterior to automatically compute 95% credible intervals for the flow estimates.

4. Results

The method was validated using the 2D Lid-Driven Cavity (LDC) problem across various Knudsen numbers ($Kn = 0.05, 0.1, 0.5, 5$) and Mach numbers.

• Proof-of-Concept (Full-Domain): The MMS-Sparse method produced velocity and shear stress estimates in high agreement with benchmark DSMC data. It provided a significant computational speedup (approx. 1.6x) compared to running a full-domain benchmark DSMC.
• Pseudo-Hybrid Approach (Near-Wall Only): To simulate a true hybrid environment, the authors used DSMC data only within $2\lambda$ of the walls. The results showed that while velocity estimates remained highly accurate, the shear stress showed some localized deviations near corners and midpoints—likely due to the singular nature of the LDC corners and the imposition of zero corrections in the interior.
• Noise Reduction: The SSBL approach successfully filtered high-frequency statistical noise from the DSMC data, producing smooth, physically consistent corrections for the CFD solver.

5. Significance

This work represents a significant step toward a "general-purpose" hybrid solver. By replacing rigid global basis functions with flexible, multilevel RBFs, the authors have bridged the gap between the mathematical rigor of Bayesian inference and the practical necessity of simulating complex, real-world geometries. This framework provides a robust pathway for simulating multi-scale flows in aerospace (re-entry) and microfluidics (MEMS) with reduced computational cost and increased reliability.