A hierarchical sparse-grid particle method for the Vlasov--Poisson system

What the math gives to ML

The transferable asset is a Galerkin interpretation of sparse-grid particle deposition: raw point samples are projected onto a hierarchical B-spline space rather than represented on a dense tensor grid. This yields a structured feature map whose parameter count scales like a sparse grid while retaining high-order approximation for functions with mixed derivative regularity. A practical neural-network transfer is a sparse-grid B-spline front end for low-dimensional continuous inputs, followed by an MLP, with resolution selected using the paper's explicit approximation-versus-sampling tradeoff. The method is most promising for coordinate networks, PINNs, neural operators, and models consuming irregular point clouds rather than high-dimensional language representations.