Learning from Linear Algebra: A Graph Neural Network Approach to Preconditioner Design for Conjugate Gradient Solvers

Abstract Large linear systems are ubiquitous in modern computational science and engineering. Their efficient solution is frequently challenging, especially in conjunction with problems described by parametric PDEs, where many such systems have to be solved. The primary approach for solving large linear systems is the use of Krylov subspace iterative methods with well-designed preconditioners. Recently, graph neural networks (GNNs) have been used in conjunction with solving parametric PDEs, and have been shown to be promising tool for designing preconditioners. Nevertheless, the GNN-based preconditioners reported in the literature have worse effect on system’s spectrum than analogous preconditioners from classical linear algebra. Here we employ well-established preconditioners from linear algebra as starting point for training GNN to obtain preconditioners that yield a more substantial reduction in the condition number of the systems when compared to classical preconditioners. Numerical experiments demonstrate the efficiency of our approach in comparison to classical and known neural network-based methods for parametric PDEs. In addition, a heuristic justification for the loss function employed in this study is provided, and we demonstrate that preconditioners obtained by learning with this loss function reduce the condition number in a way that is more desirable for the conjugate gradient iterative method.