Answers are grounded in this paper and cite their sources.
The deep energy method (DEM) provides a meshfree variational framework for solving solid mechanics problems with neural networks, but it suffers from displacement oscillations, integration errors, and convergence failures when applied to heterogeneous composites. The shallow energy method (SEM) replaces the deep multilayer perceptron of the DEM with a radial basis function neural network (RBFNN), improving conditioning but retaining the limitations of the meshfree formulation. This work proposes the mesh‐based shallow energy method (M‐SEM), which embeds the SEM approach in a finite element discretization. The mesh constrains the displacement field via shape function interpolation, replaces automatic differentiation with algebraic shape function derivatives, and enables high‐order Gaussian quadrature, while the RBFNN yields a well‐conditioned optimization landscape. The approach is validated on a coupon with a stiff inclusion through a full comparison against the DEM, the SEM, and a mesh‐based DEM, and on a fiber‐reinforced polymer microstructure across four packing configurations sharing identical hyperparameters. The M‐SEM reduces stress errors by over two orders of magnitude relative to the DEM at a fraction of the runtime, performs consistently across six radial basis function kernels, and converges robustly across all microstructure variants.
No takes yet. Share an insight, caveat, or question.
Wichmann et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: