Two-Stage Intelligent Inverse Analysis on an Irregular Deep Excavation Considering Three-Dimensional Effects

This study proposes a two-stage intelligent inverse analysis method (TSIIAM) to overcome the common limitation of soil parameter inverse analysis in deep excavations, which typically relies on a single cross section and ignores the potential impact of inherent three-dimensional (3D) effects. The intelligent inversion model is developed using deep learning algorithms and takes multiple deformation sequences of diaphragm walls along the depth as input features to identify and optimize key parameters of the hardening soil with the small-strain stiffness model and the soil–wall interface parameter. Stage I aims to identify key parameters and indicate the direction toward the approximate optimal solution, while Stage II focuses on improving inversion accuracy and rapidly approaching the optimal solution. A case study of an irregular deep excavation project is conducted to demonstrate the accuracy and reliability of the TSIIAM. The results indicate that predicted wall deformations using inversion parameters agree well with measurements across excavation levels. Compared with the initial basic soil parameters, the average weighted root mean squared error based on inversion parameters decreases in both Stage I and Stage II and exhibits slight fluctuation after reaching a lower value in Stage II. This trend reflects a progressive reduction in overall error between predicted and measured wall deformations, suggesting that the inversion model can effectively capture the implicit spatial features and spatial coupling relationships embedded in multiple deformation sequences. The proposed method achieves a balance between training cost and prediction accuracy under relatively small data sets, offering an effective pathway for multiparameter inversion of advanced constitutive models considering 3D effects in complex, irregular excavations.