Differential settlement control in highway widening projects on soft soil remains a major challenge. This study investigates the mechanisms of differential settlement in widened embankments and develops an intelligent prediction framework by integrating high-fidelity numerical simulations with physics-constrained deep learning. First, comprehensive numerical simulations were performed using a Hardening Soil (HS) model considering structural degradation in PLAXIS 2D. This work revealed the redistribution of additional stress under widening loads and elucidated the evolution mechanisms of plastic zone development and interface shear behavior at the junction of new and existing subgrades. A reasonable step width range of 1.5–2.0 m is identified based on deformation control, plastic zone extent, and engineering economy. To overcome the limited physical consistency of conventional deep learning, a physics-constrained WOA-Attention-BiGRU framework is developed by embedding monotonicity and logarithmic consolidation rate decay as regularization terms in the loss function, significantly improving robustness and interpretability in data-sparse regimes. Validation shows R2 = 0.988 and RMSE ≤ 3.2 mm for full-time-series settlement prediction, substantially outperforming pure data-driven models; slight error increase occurs under extreme soft conditions but remains within engineering tolerances. The findings elucidate the nonlinear coupling mechanisms between geometric parameters and foundation stiffness governing differential settlement gradients, providing reliable support for optimized design and long-term settlement assessment in soft soil highway widening projects.
Li et al. (Thu,) studied this question.