RBF-FD discretization of the Navier-Stokes equations on scattered but staggered nodes

Tianyi Chu, Oliver T. SchmidtJournal of Computational Physics 2023 · DOI

TL;DR

This paper introduces a novel RBF-FD (Radial Basis Function-Finite Difference) method for solving the incompressible Navier-Stokes equations on scattered, staggered grids. This approach offers a promising way to achieve accurate and stable fluid simulations with significantly reduced computational cost and grid resolution compared to existing methods, without requiring special treatments near boundaries.

What Problem Does This Paper Address?

Simulating fluid dynamics, particularly turbulent flows governed by the Navier-Stokes equations, is computationally demanding. Traditional methods often rely on structured grids, which can be inefficient for complex geometries and may suffer from numerical instabilities like odd-even decoupling. While mesh-free methods like Radial Basis Functions (RBFs) offer flexibility, global RBF methods can be computationally expensive and prone to ill-conditioning. Local RBF-FD methods have emerged as a promising alternative, combining the flexibility of mesh-free approaches with the efficiency of finite differences. However, applying these methods to the Navier-Stokes equations, especially on complex, non-uniform grids, while maintaining accuracy and stability, remains a challenge. This paper aims to address these issues by developing an RBF-FD method that leverages a staggered grid arrangement and a specific type of RBF (polyharmonic splines with polynomial augmentation) to achieve high accuracy and stability with reduced computational resources.

Methods

The core of this work is the development of a semi-implicit fractional-step algorithm that utilizes Radial Basis Function-Finite Differences (RBF-FD) on a scattered, staggered node layout.

  • RBF-FD Discretization: The paper employs polyharmonic splines (PHS) augmented with polynomials (PHS+poly) as the radial basis functions. This choice is motivated by their ability to avoid shape parameters, which are often problematic in other RBFs, and their proven accuracy for derivative approximations. The polynomial augmentation helps to improve consistency with Taylor series expansions and mitigate stagnation errors.
  • Staggered Grid: A key innovation is the use of a staggered grid arrangement. Unlike traditional collocated grids where all variables are computed at the same points, this method places pressure at the vertices of underlying triangular meshes (P-grid) and velocities at the midpoints of cell edges (V-grid). This staggering is inspired by classical Marker-And-Cell schemes and finite volume methods and is crucial for preventing numerical instabilities like odd-even decoupling.
  • Fractional-Step Method: A three-stage semi-implicit fractional-step algorithm is used to advance the solution in time. This method decouples the velocity and pressure computations, allowing for efficient time integration. The stages involve explicit advancement of convective terms, implicit treatment of viscous terms, and a pressure-correction step to enforce incompressibility.
  • Parameter Optimization: A systematic parameter study was conducted to identify optimal settings for the RBF-FD stencil size ($n$), PHS exponent ($m$), and polynomial degree ($q$). This was achieved by minimizing the truncation error for a wave-like test function on a heterogeneous grid.
  • Modified Wavenumber Analysis: To rigorously assess the accuracy of the RBF-FD discretization on scattered nodes, the authors extend the classical modified wavenumber analysis. This technique quantifies the error in representing wave propagation and is crucial for understanding the spectral accuracy of the scheme.

Datasets

The paper does not explicitly mention or use "datasets" in the machine learning sense. Instead, it relies on:

  • Synthetic Test Functions: A wave-like test function ($g(x) = \cos(k_x x + k_y y)$) is used extensively for parameter selection and error analysis.
  • Benchmark Problems: The developed solver is validated against two well-established benchmark problems in computational fluid dynamics:
  • Lid-driven cavity flow: This problem is used to test the solver's ability to handle internal flows and achieve accurate results across a range of Reynolds numbers (up to Re = 10000).
  • Flow over a cylinder: This is a classic unsteady benchmark problem used to demonstrate the solver's capability in simulating wake phenomena and vortex shedding (at Re = 100 and 200).

Key Insights and Ideas

  • Staggering is Key for Stability and Efficiency: The staggered grid, combined with RBF-FD, is shown to be highly effective in preventing numerical instabilities and allows for significantly coarser grids while maintaining accuracy. This is a major step towards more efficient simulations.
  • PHS+poly RBFs Offer Robustness: The choice of polyharmonic splines with polynomial augmentation proves to be a robust RBF choice, avoiding the need for tricky shape parameter tuning and offering good accuracy for derivative approximations.
  • Modified Wavenumber Analysis on Scattered Grids: Extending this analysis to scattered, heterogeneous grids is a significant contribution, providing a rigorous way to quantify the accuracy of RBF-FD schemes beyond simple grid refinement studies.
  • Spectral-Like Accuracy Achieved: The optimized RBF-FD scheme achieves accuracy comparable to high-order Padé-type finite differences, particularly up to a recommended maximum modified wavenumber of $k\Delta r = \sqrt{2}$. This suggests the potential for high-fidelity simulations.
  • Reduced Computational Cost: The combination of staggered grids and optimized RBF-FD stencils leads to a substantial reduction in the number of grid points and stencil size compared to previous RBF-based methods for similar problems.

Key Results

  • Optimal Parameters Identified: The optimal parameter combination for the RBF-FD scheme was found to be a stencil size of $n=28$, PHS exponent $m=7$, and polynomial degree $q=3$. This combination provides a good balance between accuracy, efficiency, and stability.
  • High Accuracy Demonstrated: The modified wavenumber analysis shows that the proposed scheme achieves accuracy comparable to 6th-order Padé-type finite differences for both first and second derivatives, even on highly heterogeneous grids. The error relative to the spectral limit is minimal ($<1.5\%$) at the recommended maximum wavenumber.
  • Stable Simulations: The solver successfully simulates the lid-driven cavity flow up to Re = 10000 and cylinder flow at Re = 100 and 200, matching benchmark results from literature. Crucially, these simulations are stable without the need for hyperviscosity or other regularization techniques.
  • Computational Savings: The staggered grid approach allows for simulations with significantly fewer nodes and smaller stencils compared to previous RBF-based methods for similar problems. For instance, the cylinder flow simulation uses approximately 65% fewer nodes on the V-grid and 85% on the P-grid compared to a related study.

Limitations and Future Work

  • Limitations:
    • The current work focuses on 2D incompressible flows.
  • While the modified wavenumber analysis is extended to scattered grids, it is still a theoretical tool for assessing discretization error.
  • The paper acknowledges that for very high wavenumbers ($k\Delta r > \sqrt{2}$), the accuracy degrades, suggesting the need for grid refinement or filtering.
  • Future Work:
    • Extension to 3D: The authors explicitly state that extending the method to three-dimensional problems is a natural next step.
  • Non-Canonical Flows: Demonstrating the solver's capability on more complex, non-canonical geophysical or engineering flows is also a future direction.
  • Strong Formulation Applications: The RBF-FD method's direct use in the strong formulation makes it suitable for constructing sparse Jacobians for hydrodynamic stability analysis, which is an area of ongoing research for the authors.
  • Optimization of Shape Parameters: While PHS+poly avoids shape parameters, for other RBFs, optimizing these parameters for specific node distributions remains an area for further investigation.

Summary updated

BibTeX
@article{chu2023rbf,
  author = {Tianyi Chu and Oliver T. Schmidt},
  title = {{RBF-FD discretization of the Navier-Stokes equations on scattered but staggered nodes}},
  journal = {Journal of Computational Physics},
  year = {2023},
  doi = {10.1016/j.jcp.2022.111756},
}
Figure