Deep Learning Accelerated Algebraic Multigrid Methods for Polytopal Discretizations of Second-Order Differential Problems

Abstract

Algebraic Multigrid (AMG) methods are state-of-the-art algebraic solvers for partial differential equations. Still, their efficiency depends heavily on the choice of suitable parameters and/or ingredients. Paradigmatic examples include the so-called strong threshold parameter $\theta$, which controls the algebraic coarse-grid hierarchy, as well as the smoother, i.e., the relaxation methods used on the fine grid to damp out high-frequency errors. In AMG, since the coarse grids are constructed algebraically (without geometric intuition), the smoother's performance is even more critical. For the linear systems stemming from polytopal discretizations, such as Polytopal Discontinuous Galerkin (PolyDG) and Virtual Element Methods (VEM), AMG sensitivity to such choices is even more critical due to the significant variability of the underlying meshes, which results in algebraic systems with different sparsity patterns. We propose a novel deep learning approach that automatically tunes the strong threshold parameter, as well as the smoother choice in AMG solvers, for linear systems of equations arising from polytopal discretizations, thereby maximizing AMG performance. We interpret the sparse matrix resulting from polytopal discretization as a grayscale image, and by applying pooling, our neural network extracts compact features that preserve the necessary information at a low computational cost. We test various differential problems in both two- and three-dimensional settings, with heterogeneous coefficients and polygonal/polyhedral meshes, and demonstrate that the proposed approach generalizes well. In practice, we demonstrate that we can reduce AMG solver time by up to $27\%$ with minimal changes to existing PolyDG and VEM codes.