Answers are grounded in this paper and cite their sources.
We present physics-informed neural network (PINN) approaches for solving elliptic boundary value problems with improved stability and accuracy. Standard PINNs based on multilayer perceptrons often exhibit limited approximation power and unstable convergence, particularly when trained with second-order optimization methods. To overcome these issues, we propose a family of localized basis-inspired architectures specifically designed to work effectively with the Levenberg–Marquardt (LM) optimizer. The Localized Basis Net (LBN) employs localized units with interpretable parameters, the extended LBN (XLBN) augments this structure with nonlinear mixing for enhanced flexibility, and the Fixed Basis Net (FBN) is a variant of XLBN where basis-related parameters are fixed to isolate and evaluate the effect of learning basis functions. We also establish the approximation capability of LBNs in Sobolev norms. Numerical experiments show that the proposed networks, when coupled with LM, achieve robust convergence and higher accuracy for a variety of elliptic problems, including linear, multiscale coefficient, and nonlinear p-Laplacian equations. These results demonstrate that network–optimizer co-design provides a promising direction for developing reliable and interpretable neural network methods in scientific computing.
No takes yet. Share an insight, caveat, or question.
Jo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: