Hybrid FEM–machine learning framework for back-analysis of spatially varying soil parameters in super-large caisson foundation

Accurate and efficient estimation of soil parameters is critical for the safe and successful construction of super-large caisson foundations, which are increasingly utilized in major infrastructure projects. Conventional in situ and laboratory methods are often slow, costly, and unable to capture dynamic soil–structure interactions during the sinking process. This study proposes a novel hybrid framework that integrates 3D finite element modeling (FEM), Uniform Design theory, and advanced machine learning (ML) for high-precision back-analysis of soil parameters. The approach is validated using the south anchorage of the super-large rectangular caisson in the Nanjing Longtan Yangtze River Bridge project. A total of 550 FEM simulations were conducted under varying soil parameter scenarios, generating corresponding stress responses. These stress–parameter pairs trained ML models to predict soil parameters from new stress data, enabling efficient back-analysis. The dataset was further augmented to 1550 samples using an ML-based synthetic data generation scheme that preserved key parameter correlations. Eighteen ML algorithms were compared; Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Target-Specific Extra Trees (TSET) achieved the highest predictive accuracy (R² ≥ 0.98), with LightGBM performing best (R² = 0.987, MAPE = 1.68%, RSR = 0.016, VAF = 98.66%). The framework successfully captured the complex nonlinear relationships between stress responses and underlying soil properties, yielding results that aligned closely with independent geotechnical investigation reports. This validated approach provides a powerful tool for the proactive failure analysis of design assumptions, offering significant practical implications for risk assessment, failure prevention, and risk mitigation in large-scale foundation engineering.