A Learnable Multigrid Framework via Graph Convolutions

1. Introduction.

• The paper formulates the target as solving a linear system Au = f arising from mesh-based PDE discretization and places its approach in the context of geometric and algebraic multigrid.Geometric multigrid uses an explicit mesh hierarchy and geometry-aware transfers; algebraic multigrid constructs coarse operators and transfers from the system matrix.Source: §1 · p. 1
• Prior work combines multigrid and neural networks through CNN-based frameworks and learned components, while some GNN approaches target unstructured grids or algebraic multigrid.The paper positions its contribution within this broader set of learning-enhanced solver approaches.Source: §1 · p. 2
• The framework embeds graph convolutions directly in the geometric multigrid cycle, making the solver a specialized neural-network architecture rather than an external surrogate.Graph convolutions are incorporated into smoothing and inter-grid transfer; the graph can be derived from the system matrix or the mesh, with solution and source values as node features.Source: §1 · p. 2
• The proposed framework aims to improve robustness and performance with lightweight parameters, low computational cost, minimal training samples, and low offline training effort.The paper investigates generalization across source terms, coefficients, geometries, and mesh sizes.Source: §1 · p. 2

2.1. Graph representation.

• For linear finite elements on a triangulations, matrix sparsity links degrees of freedom at mesh-connected vertices, inducing a graph whose nodes are vertices and whose edges follow mesh connectivity.The graph may also be built algebraically from off-diagonal nonzeros of the system matrix.Source: §2.1 · p. 3
• Geometric and algebraic graph representations can differ for vector-valued or higher-order problems: geometric graphs use fewer nodes with richer features, while algebraic graphs follow scalar system-matrix sparsity.The framework can use either geometric or algebraic multigrid once its graph, smoother, and transfer operators are fixed.Source: §2.1 · pp. 3–4

2.2. Classical geometric multigrid method.

• A geometric multigrid V-cycle smooths fine-grid errors, restricts the residual, solves a coarse-grid error equation, prolongates the correction, and post-smooths.Smoothers reduce high-frequency error; coarse grids help eliminate error components that are smooth on finer grids.Source: §2.2 · pp. 4–5
• Geometric coarsening builds coarse operators by discretizing on coarse meshes, whereas Galerkin coarsening uses A_(ℓ+1) = R_ℓ A_ℓ P_ℓ; geometric transfers need not be transposes but must preserve approximation quality.Insufficiently accurate transfers can make coarse-grid correction ineffective and cause poor convergence or divergence.Source: §2.2 · p. 5

2.3. Graph neural networks.

• Graph convolutions update node features using local information from each node and its neighbors, with the output depending on edge weights as well.For PDE solvers, this local aggregation is analogous to relaxation such as Jacobi; the right-hand side can be an additional feature.Source: §2.3 · p. 6
• The framework integrates learnable graph convolutions into the multigrid cycle, while multilevel pooling and unpooling correspond to restriction and prolongation.The stated residual-based loss is the norm of f − Au.Source: §2.3 · p. 6, §3 · p. 7

3. A learnable multigrid framework via graph convolutions.

• Graph convolutions are integrated into smoothing, restriction, and prolongation within the V-cycle, whose levels and learnable components define a GNN architecture.The method differs from hybrid approaches that apply a neural solver after the classical smoother.Source: §3 · pp. 7–8

3.1. Training as an iterative solver.

• Training is unsupervised: each epoch runs one V-cycle, minimizes the linear-system residual norm, and updates learnable parameters by backpropagation.The residual itself supplies the loss, so labeled data are not required.Source: §3.1 · p. 9

3.2. Graph convolution operators.

• The framework uses SAGEConv with mean aggregation, while allowing other graph convolutions such as ChebConv, which requires more computational resources.The paper chooses one-layer graph convolutions to keep parameter count and offline training costs low.Source: §3 · p. 8, §3.2 · p. 10

3.3. Spectral analysis of the learnable smoother.

• The learnable smoother’s minimum smoothing factor cannot exceed optimal damped Jacobi’s, and frequency-dependent damping can improve cases where uniform damping is suboptimal.The additional degree of freedom can give different eigenmodes different effective damping.Source: §3.3 · p. 11

3.4. Initialization of learnable parameters.

• Identity graph-convolution weights and classical scaling-factor values initialize the solver to behave exactly like classical multigrid at the first training iteration.For damped Jacobi, the scaling factors are initialized as θℓ=ω and ηℓ=1.Source: §3.4 · p. 12

4.1. Darcy equation.

• LMG-RP converges in about 21 epochs; piecewise constant interpolation leaves learning room where bilinear multigrid’s roughly 10-iteration convergence does not.Both learnable variants improve the piecewise constant multigrid baseline.Source: §4.1 · pp. 13–14

4.1.1. Performance and discussion.

• Both learnable solvers outperform classical multigrid on all three Darcy testing fields; their V-cycles are cheaper, and LMG-RP is faster on Fields II–III.LMG-RPS is best for Field I; the cheaper cycles result from sparser coarse operators with piecewise constant interpolation.Source: §4.1.1 · p. 14, §4.1 · p. 14, Table 1 · p. 14
• LMG-RP and LMG-RPS reduce high- and low-frequency residual components more uniformly than classical multigrid, whose low-frequency components persist.The frequency contours are consistent with the paper’s description of frequency-dependent damping.Source: §4.1.1 · p. 16, Figure 5 · p. 16

4.1.2. Sensitivity to the initialization.

• Random initialization is less stable than identity initialization and can diverge; the random model diverged on Field II, while identity initialization starts from stable classical multigrid.The tested random-initialized model converged after about 130 epochs in one run.Source: §4.1.2 · pp. 16–17

4.2. Helmholtz equation.

• Training takes about 40 epochs across Helmholtz hierarchy sizes; learnable multigrid remains robust as classical multigrid degrades, despite V-cycle inference costing around 1.5 times as much.The reported cost per iteration means fewer iterations do not reduce total computational time.Source: §4.2 · p. 18, Table 2 · p. 18

4.2.1. Performance enhancement via pre-training strategies.

• Pre-training improves FMG-preconditioned FGMRES results overall at low additional cost, though not every run improves every case.At J = 5, the learnable solver outperforms the classical solver only after pre-training.Source: Table 4 · p. 19, §4.2.1 · p. 20

4.2.2. Zero-shot super-resolution via weight transfer.

• Without fine-tuning, transferred weights at k = 32π yield a learnable preconditioner that converges in all reported cases and significantly outperforms the classical one.The best reported result occurs at J = 8.Source: §4.2.2 · p. 20, §4.2.1 · p. 20, Table 5 · p. 21

4.2.3. The partial weight transfer strategy for generalization to different wave numbers, geometries, and the number of levels.

• Partial transfer improves iteration counts and wall-clock time for all tested J; nearly constant iteration counts indicate effective handling of level mismatch.Matching smoother levels receive learned weights, mismatched levels use the classical smoother, and extra coarse levels copy compatible weights.Source: §4.2.3 · p. 21

• The solver adds only a few learnable parameters through graph convolutions embedded in multigrid components, retaining classical multigrid interpretability and theoretical foundation.Source: §5 · pp. 21–22
• Experiments on heterogeneous Darcy and impedance Helmholtz equations show generalization across conductivity fields and mesh sizes, and improved convergence over classical multigrid for Helmholtz.Source: §5 · p. 22
• Pre-training and weight transfer support larger-scale problems and higher wave numbers with reduced training costs; Tables 6–7 describe partial transfer with an FMG preconditioner.Source: §5 · p. 22, Table 6 · p. 22, Table 7 · p. 22
• Future work includes PDE-tailored graph convolutions, adaptive smoothing-step selection, and multi-stage pre-training for broader generalization.Source: §5 · p. 22