Physics-Informed Neural Networks (PINNs) provide a promising framework for solving partial differential equations (PDEs). By incorporating temporal causality, Causal PINN improves training stability in time-dependent problems. However, applying Causal PINN to higher-order nonlinear PDEs, such as the Cahn–Hilliard equation (CHE), presents notable challenges due to the inefficient utilization of temporal information. This inefficiency often results in numerical instabilities and physically inconsistent solutions. This study systematically analyzes the limitations of Causal PINN in solving the one-dimensional CHE. To resolve these issues, we propose a novel framework called APM (Adaptive Progressive Marching)-PINN that enhances temporal representation and improves model robustness. APM-PINN mainly integrates a progressive temporal marching strategy, a causality-based adaptive sampling algorithm, and a residual-based adaptive loss weighting mechanism (effective with the chemical potential reformulation). Comparative experiments on two one-dimensional CHE test cases show that APM-PINN achieves relative errors consistently near 10−3 or even 10−4. It also preserves mass conservation and energy dissipation better. The promising results highlight APM-PINN’s potential for the accurate, stable modeling of complex high-order dynamic systems.
Hu et al. (Mon,) studied this question.