Dynamic Divergence-Free Kernels

• Dynamic Divergence-Free Kernels (DDFKs) are mesh-free, memory-efficient representations that enforce the incompressibility constraint through analytically divergence-free kernels.
• They combine radial basis function methods with kernel-based techniques to bypass pressure projection and minimize numerical dissipation in complex fluid flows.
• DDFKs utilize adaptive kernel centers, efficient neighbor searches, and GPU parallelization to achieve high fidelity in time-dependent simulations of vortex-dominated flows.

Dynamic Divergence-Free Kernels (DDFKs) are mesh-free, memory-efficient representations for incompressible fluid simulation. By constructing the velocity field as a sum of compactly supported, analytically divergence-free kernels, DDFKs rigorously embed the incompressibility constraint, enabling accurate and robust simulation of complex, time-dependent flows with minimal numerical dissipation. DDFKs combine the strengths of radial basis function methods and kernel-based representations with tailored optimization algorithms, providing a viable alternative to both grid-based discretizations and neural implicit approaches for representing and evolving incompressible velocity fields (Xing et al., 25 Jan 2026, Ni et al., 2 Apr 2025).

1. Mathematical Foundation: Single Divergence-Free Kernel

Each DDFK is built from a matrix-valued, compactly supported radial basis function (RBF) and a local vector circulation weight. For spatial dimension $d$ and kernel index $i$, the key components are:

• Center $p_i \in \mathbb{R}^d$
• Support radius $h_i > 0$
• Weight $w_i \in \mathbb{R}^d$
• RBF: The Wendland C⁴ RBF,

$$\phi(r) = (1 - r)^6 (35 r^2 + 18 r + 3), \text{ for } 0 \leq r \leq 1,$$

zero otherwise.

The $i$-th divergence-free kernel basis at point $x$ is

$$\psi_i(x) = (-I \Delta + \nabla\nabla^T)[\phi(r_i)], \quad r_i = \|x - p_i\| / h_i,$$

where $I$ is the identity and derivatives are with respect to $x$. The induced vector field (single DFK) is

$\tilde{u}_i(x) = \psi_i(x) w_i.$

By vector calculus identities, $\nabla \cdot \tilde{u}_i(x) = 0$ pointwise for any $w_i$, ensuring incompressibility at the kernel level. The kernels admit explicit closed forms as combinations of $f(r)$ and $g(r)$ functions:

$\psi_i(x) w_i = f(r) w_i + g(r) (w_i \cdot y) y,$

where $y = (x-p_i)/h_i$, $f(r) = 56 (1 - r)^4 [-5(d+5) r^2 + (d-1)(4r+1)]$, $g(r) = 1680 (1 - r)^4$ for $0 \leq r \leq 1$ (Xing et al., 25 Jan 2026, Ni et al., 2 Apr 2025).

2. Global Velocity Representation and Incompressibility

The full time-dependent velocity field is represented as a sum of $N$ dynamically evolving DDFKs:

$$u(x, t) = \sum_{i=1}^N \psi_i(x; p_i(t), h_i) w_i(t)$$

Each center $p_i(t)$ evolves according to the local flow (advection), while $w_i(t)$ controls local circulation and detail. Optionally, $h_i$ can be adapted to resolve multiple length scales (e.g., large eddies with large $h_i$, fine-scale vortices with small $h_i$).

Because every $\psi_i$ is analytically divergence-free and linear in $w_i$, $u(x, t)$ remains divergence-free everywhere, eliminating the need for pressure projection, Poisson solves, or additional incompressibility penalties at the PDE-discretization level (Xing et al., 25 Jan 2026, Ni et al., 2 Apr 2025).

3. Temporal Evolution and Solver Algorithm

The DDFK simulation advances through three central phases at each time step:

1. Initialization: At $t=0$, fit $\{p_i(0), w_i(0)\}$ to an initial velocity $u^0(x)$ by minimizing

$$L_{\text{init}} = \mathbb{E}_x[\|\tilde{u}^0(x) - u^0(x)\|_1] + \lambda_{\text{grad}} \mathbb{E}_x[\|\nabla \tilde{u}^0(x) - \nabla u^0(x)\|_1].$$

1. Time-Stepping:
   • Reinitialization (every $N_{\text{reinit}}$ frames): Regenerate the kernel set to avoid clustering or voids.
   • Advection of Centers: Centers $p_i$ are advected using fourth-order Runge-Kutta (RK4) with the current velocity field.
   • Optimization: Starting from advected positions and current weights, solve a physics-based optimization:

   $$L_{\text{adv}} = L_{\text{vor}} + \lambda_{b1} L_{b1} + \lambda_{b2} L_{b2}$$

   where $L_{\text{vor}}$ enforces the vorticity evolution from the (inviscid) Navier-Stokes equations, $L_{b1}$ imposes no-slip on solid boundaries $\Gamma_1$, and $L_{b2}$ imposes free-slip on other boundaries $\Gamma_2$.

2. Acceptance: Update $\{p_i, w_i\}$ with the optimized parameters.

By construction, DDFK solvers are first-order in time, trading some efficiency for superior treatment of incompressibility and reduced numerical dissipation, making them particularly suitable for vortex-heavy or long-time simulations where accurate preservation of fine-scale features is critical (Xing et al., 25 Jan 2026).

4. Implementation Aspects and Computational Complexity

• Cost per Query: Naïvely $O(N)$ per velocity evaluation, but practical implementations reduce this to $O(1)$ via hash-based local neighbor lists, akin to neighbor searches in particle methods.
• Memory Efficiency: Stores $N$ kernel centers, radii, and weights, so $O(Nd)$ storage, with adaptive spatial support.
• Differentiation: All required derivatives (gradients, vorticity, etc.) admit closed-form expressions, making the approach far cheaper than auto-differentiation through deep neural networks.
• Kernel Re-initialization: To combat particle collapse or void formation, periodic re-fitting of kernels is used, with additional kernels inserted in high-error regions as needed.
• Parallelization: Both the kernel evaluations and neighbor search can be efficiently batched and parallelized on GPU architectures (Xing et al., 25 Jan 2026).

5. Quantitative and Qualitative Performance

DDFKs demonstrate robust and accurate performance across standard benchmarks, with comparisons to neural representation (INSR), Gaussian spatial representation (GSR), and conventional grid-based solvers. In the Taylor-Green vortex scenario on a $60\times60$ grid, the mean-squared error (MSE) over time is as follows:

Method	Frame 0	Frame 50	Frame 100
Semi-imp Euler	$2.43\times10^{-4}$	$9.76\times10^{-3}$	$2.02\times10^{-2}$
INSR	$8.99\times10^{-7}$	$1.72\times10^{-5}$	$1.99\times10^{-5}$
NMC	$1.83\times10^{-4}$	$6.49\times10^{-4}$	$1.73\times10^{-3}$
GSR	$9.96\times10^{-8}$	$2.51\times10^{-7}$	$2.18\times10^{-7}$
DDFK	$2.34\times10^{-8}$	$2.48\times10^{-8}$	$3.31\times10^{-8}$

Key evaluation criteria include vortex and flow structure preservation, exact zero divergence ($\|\nabla\cdot u\|_{L^2} \equiv 0$), and reduced numerical dissipation over long times. Qualitative results include:

• Precise maintenance of separation in leapfrog vortex tests (where GSR merges vortices prematurely).
• Sharp wakes and higher effective Reynolds numbers in Kármán vortex street simulations, without spurious noise.
• Exact preservation of flowrates across boundaries in piston-in-cylinder setups, unlike methods that exhibit leakage (Xing et al., 25 Jan 2026).

6. Comparison to Other Divergence-Free Kernel Approaches

DDFKs are situated among several advanced divergence-free kernel approaches:

• Composite B-spline (CBS) Kernels: CBS kernels, as used in immersed finite element-difference (IFED) methods, offer discrete-to-continuous transfer operators that maintain divergence-free interpolation at the Lagrangian level. CBS kernels yield superior volume conservation and can converge on coarser grids versus classical IB or standard B-spline kernels (Li et al., 2024).
• Divergence-Free Interpolated Boundary (DFIB) Method: The DFIB approach constructs interpolation and force-spreading operators to guarantee continuous divergence-free velocity, C¹ smoothness, and adjoint energy transfer. While DDFKs emphasize mesh-free, adaptive representations for general PDEs, DFIB is designed for staggered-grid fluid-structure interaction and supports smooth force transfer with vastly improved volume (area) conservation, albeit with somewhat higher computational cost (Bao et al., 2017).
• Neural Representations: Implicit neural spatial representations (INSRs) can offer similar adaptive expressivity but typically require explicit $\nabla \cdot u$ penalties or projection, with higher parameter count and increased cost for differentiation (Xing et al., 25 Jan 2026, Ni et al., 2 Apr 2025).

7. Strengths, Limitations, and Ongoing Developments

DDFKs provide several notable strengths:

• Exact Incompressibility: The pointwise divergence-free condition is hardwired into the representation.
• Vorticity and Feature Retention: Kernels allow for multiscale, high-fidelity preservation of vortex cores and fine details over long times.
• Reduced Numerical Dissipation: Lower discretization error leads to minimal artificial damping.
• Memory and Adaptivity: Mesh-free and compact, with local support for efficient evaluation.

Current limitations:

• Time Integration Order: The present solver is first-order in time, which can be slower relative to highly optimized grid-based methods.
• Viscosity and Higher-Order Physics: Extensions to implicit/explicit handling of viscosity, higher-order time-stepping, and hybridization with grid methods are active areas for future work.
• Optimization Cost: Each step requires local or global optimization, which, despite batching and GPU acceleration, may lag behind specialized solvers for structured problems.

Ongoing research is directed towards higher-order integrators, further adaptivity in kernel parameters, coupling to viscous terms, and hybrid solvers that combine kernel and grid-based representations (Xing et al., 25 Jan 2026, Ni et al., 2 Apr 2025).

---

References:

• J. Xing et al., "DDFKs: Fluid Simulation with Dynamic Divergence-Free Kernels" (Xing et al., 25 Jan 2026)
• X. Ni et al., "Representing Flow Fields with Divergence-Free Kernels for Reconstruction" (Ni et al., 2 Apr 2025)
• L. Li et al., "Local Divergence-Free Immersed Finite Element-Difference Method Using Composite B-Splines" (Li et al., 2024)
• B. Bao et al., "An Immersed Boundary Method with Divergence-Free Velocity Interpolation and Force Spreading" (Bao et al., 2017)