Kernelized Stable Fluids for simulating physically accurate solutions
math.NA, cs.NA
Submitted: 2026-11-02
Updated: 2026-11-08
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
The gist: The Stable Fluids algorithm is a fractional step method for solving the incompressible Navier-Stokes (NS) equations in an unconditionally stable way.
Abstract
The Stable Fluids algorithm is a fractional step method for solving the incompressible Navier-Stokes (NS) equations in an unconditionally stable way. It was initially developed for visual effects and allows real-time simulations because of its low computational complexity. The method solves the advection step by using the method of characteristics which is implemented via linear interpolation. This results in excessive diffusion of vorticity and is one of the main reasons for the physical inaccuracy of the algorithm. We propose Kernelized Stable Fluids with Sparse Cholesky Preconditioning (KSF-Chol) in order to address this shortcoming and make the algorithm thereby suitable for engineering problems and scientific computing. The algorithm is modified by replacing the linear interpolation approach with kernel interpolation in the advection step. We accelerate the computationally expensive kernel interpolation by using sparse Cholesky preconditioning and a near-/far-field approach and optimize both methods for computations on a uniform Cartesian grid. KSF-Chol achieves per time step a near-linear computational complexity of O(N d(N/epsilon chol)) in space and time. We test KSF-Chol against reference solvers and numerical reference data on multiple 2D and 3D benchmarks and demonstrate a significantly reduced numerical diffusion as well as a competitive accuracy concerning the computation of relevant flow quantities.
Sources
- KROM: Kernelized Reduced Order Modeling
- Optimized M2L Kernels for the Chebyshev Interpolation based Fast Multipole Method
Related papers
- Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
- A Neural-preconditioned Poisson Solver for Mixed Dirichlet and Neumann Boundary Conditions
- Advanced Linear Algebra with Applications - Part I (Numerical linear algebra for PDEs, machine learning, and data assimilation)
- Operator learning meets inverse problems: A probabilistic perspective
- Second-order consistency for learning chaotic dynamics via randomized Jacobian matching
- Windowed thinning and query complexity for the bouncy particle and Zigzag samplers