Abstract
Solving partial differential equations (PDEs) using neural networks hasbecome a central focus in scientific machine learning. Training neural networksfor singularly perturbed problems is particularly challenging due to certainparameters in the PDEs that introduce near-singularities in the loss function.In this study, we overcome this challenge by introducing a novel method basedon homotopy dynamics to effectively manipulate these parameters. From atheoretical perspective, we analyze the effects of these parameters on trainingdifficulty in these singularly perturbed problems and establish the convergenceof the proposed homotopy dynamics method. Experimentally, we demonstrate thatour approach significantly accelerates convergence and improves the accuracy ofthese singularly perturbed problems. These findings present an efficientoptimization strategy leveraging homotopy dynamics, offering a robust frameworkto extend the applicability of neural networks for solving singularly perturbeddifferential equations.