LDRD
Graph Neural Networks-based Coarse Spaces for Scalable Robust Multilevel Schwarz Domain Decomposition Preconditioners
For the robustness and scalability of two-level domain decomposition methods, one of the critical components is the coarse space. Various methods, such as spectral coarse spaces, have been proposed to enhance the coarse space functions for heterogeneous problems. Though they are robust, the computational overheads associate with such approach is often rather high. In this talk, we investigate the potential of graph neural networks to predict effective coarse space functions, using only the algebraic information available through the local subdomain matrices.