Prediction and analysis of displacement response of submerged floating tunnel under impact loading based on machine learning and SHAP methods

Submerged floating tunnels (SFTs) are prone to accidental impact hazards such as ship anchoring and vessel collisions. Existing displacement prediction methods suffer from limited working condition coverage, ambiguous nonlinear multi-factor coupling mechanisms, and the black-box limitation of traditional machine learning (ML). This work develops an interpretable high-precision prediction framework for impact-induced SFT tube displacement responses. A comprehensive dataset is constructed by combining drop-hammer impact model tests and high-fidelity ABAQUS fluid-structure interaction simulations covering multiple impact, structural and marine environmental control parameters. Six mainstream ML algorithms including AdaBoost, GBDT, XGBoost, Random Forest, MLP and SVR are systematically optimized via grid search and 5-fold cross-validation. The optimized XGBoost model achieves optimal generalization performance with a test set coefficient of determination R² = 0.956. Game-theory-based SHAP interpretability analysis quantitatively ranks the dominant influencing factors and clarifies their monotonic action laws: the primary control factor exerts a displacement amplification effect, while the core structural parameter presents a significant inhibitory effect on structural deformation. This study establishes an interpretable data-driven prediction method and develops a supporting interactive prediction program, which can provide direct technical reference for anti-impact design and safety evaluation of submerged floating tunnels.