Answers are grounded in this paper and cite their sources.
Finite element analysis (FEA) faces significant challenges in maintaining mesh quality for unstructured Delaunay meshes with complex geometric features, whereas conventional optimization methods often encounter low-quality local optima. This paper proposes a novel deep deterministic policy gradient (DDPG)-driven framework to autonomously optimize node positions. First, nodes are initialized at the centroids of local polygon regions (composed of connected triangles around the target nodes) to establish topologically valid configurations; then, DDPG optimizes positions within these constrained regions. The DDPG reward mechanism integrates a distance penalty with element quality measurements, resulting in a balanced approach that outperforms methods that rely solely on quality-based rewards. This strategy enhances the overall mesh quality, improves the worst-element quality, and effectively eliminates low-quality elements while achieving strong results with small training datasets. Extensive experiments on 2D domains with curved boundaries and irregular internal voids (e.g., hollow mechanical components) demonstrate superiority over conventional methods such as Laplacian smoothing, centroidal Voronoi tessellation, and optimal Delaunay triangulation. Crucially, compared with a reference calculation solution in a load-bearing case study, optimized meshes substantially reduce FEA numerical errors—exemplified by stress calculation errors decreasing from 11.77% to 8.49%. Furthermore, DDPG-optimized meshes yield higher accuracy at lower densities and accelerate solution convergence with respect to the mesh density. This work introduces a deep reinforcement learning-based adaptive optimization framework for complex Delaunay triangulations, providing a robust tool for high-fidelity simulations.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: