Vertex-based Graph Neural Solver and its Application to Linear Elasticity Equations

Yun Liu, Chen Cui, Shi Shu, Zhen WangarXiv 2025 · DOI · arXiv

Abstract

This paper introduces a hybrid iterative solver for linear elasticity equations that combines traditional numerical methods with deep learning. By integrating a weighted block Jacobi smoother with a novel Multilevel Adaptive Graph Fourier Neural Solver (ML-AG-FNS), the authors achieve mesh-independent convergence and significantly outperform state-of-the-art algebraic multigrid (AMG) methods, particularly in complex 3D anisotropic scenarios.

Introduction: The Challenge of Linear Elasticity

In computational mechanics, solving linear elasticity equations—which describe how solids deform under load—is a fundamental task. However, when dealing with "real-world" complexity like unstructured meshes, material heterogeneity (varying stiffness), and strong anisotropy (properties changing with direction), traditional iterative solvers often stall.

The core issue is a "spectral disparity." Standard smoothers, such as the Jacobi method, are effective at dampening high-frequency error components but struggle with low-frequency, global errors. Furthermore, in near-incompressible materials, traditional methods fail to dampen "shear" modes effectively. While Multigrid methods attempt to address this by using a hierarchy of grids, designing effective "coarse-grid" operators for complex, vector-valued PDEs on irregular meshes remains a significant manual and heuristic challenge.

Methodology: The ML-AG-FNS Approach

The authors propose a Neural Solver approach that follows a "Smoothing + Correction" hybrid paradigm:

  1. Smoothing Step: Employs a traditional weighted block Jacobi method to quickly eliminate high-frequency errors.
  2. Global Correction Step: Utilizes a Graph Neural Network (GNN) based Fourier solver to handle the remaining low-frequency errors.

Key Innovations

  • Graph Fourier Neural Solver (G-FNS): Unlike standard Fourier Neural Operators (FNOs) that require regular grids, G-FNS uses GNNs to map the problem onto a graph structure, making it compatible with arbitrary unstructured meshes.
  • Adaptive Coordinate Transformation: The model learns a mapping from physical coordinates to a latent space. This allows the Fourier transform to use a "dynamic, problem-dependent basis" that captures multi-scale features better than a fixed basis.
  • Multilevel Hierarchy (ML-AG-FNS): To handle 3D systems, the authors developed a multilevel architecture that corrects errors across cascading frequency bandwidths (from fine to coarse), similar to the logic of multigrid but executed in the frequency domain.
  • Theoretical Rigor: The authors provide a formal proof of unconditional convergence using the energy norm and Korn’s inequality, ensuring the solver remains stable regardless of mesh refinement.

Key Insights and Ideas

Deep neural networks naturally suffer from "spectral bias," meaning they learn low frequencies much faster than high ones. Instead of fighting this, the authors use it to their advantage, letting the neural network handle the global low-frequency modes that traditional smoothers miss.

Spectral analysis of the error dampening factor across frequency components $\theta_1$ and $\theta_2$ for different solver configurations.

By introducing learnable coordinate transformations, the solver can "warp" the space to simplify the underlying physics, allowing the Fourier-based correction to be much more effective in heterogeneous media. Mapping the linear system to a graph (nodes for physical points, edges for stiffness matrix blocks) allows the GNN to naturally preserve the vector coupling between displacement components across x, y, and z directions.

Experimental Results

The researchers evaluated their method on four distinct categories of synthetic datasets generated via Finite Element Method (FEM) simulations, ranging from 2D isotropic materials to complex 3D orthotropic anisotropic cases. The proposed ML-AG-FNS demonstrated superior performance compared to the standard Smoothed Aggregation Algebraic Multigrid (SA-AMG):

  • Significant Speedup: In 3D orthotropic anisotropic cases, the neural solver required only ~8.5 iterations to converge, while the best SA-AMG variant required ~238 iterations.
  • Robustness: As a preconditioner for Krylov subspace methods (like FGMRES), the neural solver reduced iteration counts by nearly 5x compared to basic G-FNS.
  • Mesh Independence: Numerical experiments confirmed the theoretical proof—convergence rates remained stable even as the number of degrees of freedom (DOFs) increased.

Conclusion and Future Work

While the ML-AG-FNS shows immense promise for linear elasticity, future work will focus on extending this to nonlinear mechanics (such as hyperelasticity or plasticity). Additionally, moving from thousands of DOFs to industrial-scale problems with millions of DOFs will require optimized, distributed parallel implementations of the graph neural operators. Achieving "zero-shot" generalization across entirely different geometric topologies remains a high-level goal for the next generation of neural solvers.

Summary updated

BibTeX
@misc{liu2025vertex,
  author = {Yun Liu and Chen Cui and Shi Shu and Zhen Wang},
  title = {{Vertex-based Graph Neural Solver and its Application to Linear Elasticity Equations}},
  year = {2025},
  eprint = {2511.21491},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.21491},
}
Figure