Year: 2026
Type: preprint
Abstract: We develop a singular-layer enhanced Finite Volume Neural Network (sl-FVNN) methodology for solving singularly perturbed boundary value problems. The proposed framework combines flux-conserving finite volume residuals with analytically derived boundary layer correctors, enabling stable and efficient approximation of sharp boundary variations induced by small perturbation parameters. Numerical experiments over a broad range of small perturbation parameters ($10^{-4}$ to $10^{-12}$), demonstrate that the sl-FVNN yields uniformly accurate predicted solutions and significantly improved convergence behavior compared to conventional FVNN methods. Further comparisons with sl-PINNs from (Gie, Hong, Jung, Appl. Anal. 2024) show that our sl-FVNNs attain comparable accuracy for reaction--diffusion and Burgers equations, while achieving significantly improved convergence for convection-dominated nonlinear problems. These results highlight the effectiveness of incorporating asymptotic layer structures into finite volume based neural networks, offering a practical and scalable framework for multiscale PDE learning in stiff regimes.
Source: arXiv (Cornell University)
Authors: Gie, Gung-Min, Kim, Junghwa
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236080133
