Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.25484

KernelPack-JAX: a meshless differentiable GPU framework for node generation and forward and inverse PDE problems

Alberto Cattaneo·M. Keith Ballard·Robert M. Kirby·Grady B. Wright·Varun Shankar
View original source ↗Read source document ↗Export metadata ↓

Abstract

We present KernelPack-JAX, a meshless, differentiable, GPU framework in JAX for solving forward and inverse PDE problems in domains whose boundaries, including internal obstacles, are given only by samples of their smooth pieces. Our node generator is based on a new stacked representation of piecewise-smooth boundaries, in which a parametric partition of unity of each smooth piece induces the supports of a curl-free implicit partition of unity of the whole boundary. This representation keeps the edges and corners of the boundary and classifies every point of space without a global fit. Boundary and interior nodes are then selected by a parallel greedy independent-set algorithm that reproduces sequential Poisson-disk sampling exactly. We discretize elliptic and parabolic problems with radial basis function-generated finite differences (RBF-FD) on stencils selected by a new visibility heuristic to avoid crossing the boundary, with ghost nodes for the boundary conditions. In addition, our framework exposes a GPU sparse direct solver to JAX as a differentiable linear solve, so that Jacobians of time-dependent problems cost a small multiple of a forward solve. We derive error estimates that separate consistency from stability and bound the parameter error by the forward error. We demonstrate high-order convergence for Darcy flow, reaction-diffusion, and advection-diffusion in 2D and 3D domains, including one with an internal obstacle, and show that the recovered parameters largely follow the forward errors. We also present timings for every stage.

Explore related subjects

math.NAcs.MScs.NA

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Alberto Cattaneo, M. Keith Ballard, Robert M. Kirby, Grady B. Wright, Varun Shankar. 2026-10-18. KernelPack-JAX: a meshless differentiable GPU framework for node generation and forward and inverse PDE problems. https://arxiv.org/abs/2610.25484

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Perturbed saddle-point problems in $\mathrm{L}^p$ with non-regular loads

In this work, we develop the continuous and discrete solvability analysis for perturbed saddle-point problems in Banach spaces with forcing terms regularised by means of a projector constructed using the adjoint of a weighted Clément quasi-interpolation. We take as driving example the linearised Poisson--Boltzmann equation (an advection-diffusion-reaction problem) in mixed form. We use perturbation arguments on the continuous and discrete levels and then derive a priori estimates that remain valid when the load that appears on the right-hand side of the "second" equation is in $\mathrm{H}^{-1}$. Further, we show a supercloseness result and {analyse convergence} of an adequate adaptation of Stenberg postprocessing for mixed advection equations with non-regular data. We provide numerical results that illustrate the convergence of the proposed scheme.

math.NA↗
arXiv2026

Born-Preconditioned Learning for Iterative Neural PDE Solvers

Simulating high-frequency waves in spatially varying media remains challenging for both classical iterative methods and neural PDE solvers. We propose Neural Preconditioned Born Series (NPBS), a framework for learning iterative neural PDE solvers. Our guiding idea is to use a simple reference preconditioner to help learn a more effective neural preconditioner. Building on the Convergent Born Series (CBS), NPBS feeds a preconditioned residual to a neural network that predicts a solution update. We train the network to reduce the preconditioned residual after each update. The same reference preconditioner is used for the network input and the loss, giving a metric-matched training objective. Experiments on heterogeneous Helmholtz problems show reduced iteration counts across neural architectures and larger gains at higher wavenumbers. Within a Krylov solver, NPBS achieves the lowest wall-clock time in the evaluated solver comparison. The framework also extends naturally to three-dimensional problems. The same metric-matched formulation also improves convergence on convection-diffusion-reaction systems and Newton linear systems for nonlinear PDEs, indicating that NPBS provides a general framework for learning neural preconditioners.

math.NA↗
arXiv2026

Flexible GMRES converges in two phases

We derive a sharp upper bound on the residuals produced by the flexible GMRES (FGMRES) method. The bound shows that FGMRES exhibits two phases of convergence depending on the residual tolerance of the inner preconditioner. For small tolerances, the convergence of FGMRES is practically geometric with a constant rate throughout, while for looser tolerances the two-phase behavior becomes more pronounced. We also show that the derived bound cannot be improved and construct an example for which it becomes an equality.

math.NA↗