Deep Learning for Singular PDEs: A Weighted Neural Network Approach
Abstract
We propose a deep learning framework for the numerical approximation of weakly singular semilinear elliptic equations subject to homogeneous Dirichlet boundary conditions and positivity constraints. These problems are challenging because the nonlinear source term becomes unbounded as the positive solution approaches zero near the boundary. The proposed approach combines three components. First, the Dirichlet boundary conditions are embedded directly into the neural representation through a hard-constraint formulation. Second, a Softplus transformation preserves positivity in the interior of the computational domain, ensuring that the singular nonlinear term remains well-defined during training. Third, a singularity-aware weighted residual loss emphasizes regions where the predicted solution becomes small and the singular behavior is more pronounced. The method is evaluated in the weak singularity regime through numerical reference solutions and controlled verification tests. Comparisons with the standard residual formulation are used to assess approximation accuracy and convergence behavior. The numerical experiments illustrate the potential of combining hard boundary enforcement, positivity preservation, and singularity-aware residual weighting for weakly singular elliptic problems.
Related
- Singular perturbations and hierarchical learning in two-layer neural networks
- An adaptive weighted extrapolation neural network for dispersive shock waves generated by square-root initial condition of the KdV equation
- The Geometry of Manifold Learning: Singular foliations via Deep Learning models
- Safety-Certified Hierarchical Control of DC Microgrids via Continuous Deep Reinforcement Learning and Covariance-Weighted Sliding Mode Synthesis
- Soliton Profiles in 1D: Enhanced Neural-Network Solvers and Constrained Deep Ritz Methods
BibTeX
@misc{oulgiht2026deep,
title = {{Deep Learning for Singular PDEs: A Weighted Neural Network Approach}},
author = {Badr Oulgiht},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.19335},
eprint = {2609.19335},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.19335},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Deep Learning for Singular PDEs: A Weighted Neural Network Approach
AU - Badr Oulgiht
PY - 2026
DA - 2026/09/16
PB - arXiv
DO - 10.48550/arxiv.2609.19335
UR - https://arxiv.org/abs/2609.19335
AB - We propose a deep learning framework for the numerical approximation of weakly singular semilinear elliptic equations subject to homogeneous Dirichlet boundary conditions and positivity constraints. These problems are challenging because the nonlinear source term becomes unbounded as the positive solution approaches zero near the boundary. The proposed approach combines three components. First, the Dirichlet boundary conditions are embedded directly into the neural representation through a hard-constraint formulation. Second, a Softplus transformation preserves positivity in the interior of the computational domain, ensuring that the singular nonlinear term remains well-defined during training. Third, a singularity-aware weighted residual loss emphasizes regions where the predicted solution becomes small and the singular behavior is more pronounced. The method is evaluated in the weak singularity regime through numerical reference solutions and controlled verification tests. Comparisons with the standard residual formulation are used to assess approximation accuracy and convergence behavior. The numerical experiments illustrate the potential of combining hard boundary enforcement, positivity preservation, and singularity-aware residual weighting for weakly singular elliptic problems.
ER -