Interface-aware adaptive layer-wise physics-informed neural networks for multilayer large-strain consolidation: a benchmark of architectures, training strategies, and trade-offs

Multilayer large-strain consolidation in stratified soils involves nonlinear pore-pressure dissipation, evolving deformation, and material discontinuities, which pose significant challenges for physics-informed neural networks (PINNs). This study develops a layer-wise adaptive PINN framework in which soil layers are represented by neural subnetworks coupled through pore-pressure and Darcy-flux continuity conditions. A controlled three-layer benchmark is used to evaluate MLP and KAN architectures, adaptive activation, adaptive weighting, and parameter sharing over three independent random seeds. The framework is further extended to four- and five-layer profiles using Shared MLP, layer-wise MLP, and KAN. Results demonstrate that no single architecture is optimal across all physical responses: parameter sharing improves mean global and settlement accuracy, KAN-based models better control selected local errors, and adaptive strategies alter pressure, settlement, and interface behavior in a metric-dependent manner. The successful four- and five-layer calculations further support the applicability of the same physical formulation as the number of internal interfaces increases, although the preferred architecture changes with the layer arrangement. Overall, pore-pressure accuracy, settlement prediction, interface consistency, and computational cost involve distinct trade-offs, so model selection should reflect the target physical quantity and profile complexity.