Year: 2026
Type: preprint
Abstract: Volumetric locking remains a fundamental challenge for neural-network Galerkin methods in nearly incompressible elasticity. We introduce the \emph{locking-free enriched linearized neural network} (LF-ENN), to the best of our knowledge, the first neural-network Galerkin method with a rigorous locking-free guarantee. Building on the enriched Galerkin framework, the method combines neural approximation with a structure-preserving enrichment that reduces the discrete incompressibility constraint to a graph-Laplacian problem. This construction enables divergence-free corrections and error estimates uniform in the Lamé parameter $λ\in(0,\infty]$. Nitsche boundary enforcement and tailored stabilization overcome the lack of conventional finite element trace inequalities for neural trial spaces. Under suitable approximation and mesh-resolution assumptions, we establish stability and error estimates independent of network width, mesh size, and $λ$, including an $O(h^s)$ estimate for reduced-regularity solutions with $s>1/2$. Numerical experiments in two and three dimensions confirm locking-free behavior and demonstrate substantial gains in displacement accuracy. In particular, LF-ENN achieves approximately twelve times smaller displacement error than a higher-order locking-free enriched Galerkin method while using fewer than one-third as many unknowns. These results establish a rigorous neural Galerkin framework that combines incompressibility robustness with markedly improved approximation accuracy.
Source: arXiv (Cornell University)
Authors: Kim, Bonyong, Lee, Young Ju
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236264636
