<p>Accurate prediction of time‑dependent multi‑responses of surrounding rock during tunnel excavation is critical for engineering safety and rational support design. Traditional numerical simulations are accurate but computationally prohibitive for real‑time use, whereas purely data‑driven models lack physical consistency. To address these issues, a physics‑informed surrogate model integrating Kolmogorov-Arnold Networks (KAN) and Long Short-Term Memory (LSTM) networks is proposed. The model takes seven static mechanical parameters of the rock mass as input and predicts the temporal evolution of deformation, stress, and damage characteristic zones including loosened, plastic, and disturbed zones induced by excavation. To overcome the lack of unified identification criteria, a quantitative determination method based on thermodynamic internal variable theory is introduced. The loosened zone is delineated by plastic shear strain, plastic volumetric strain, and their depth‑correlated Pearson coefficient curves; the disturbed zone boundary is determined using the cumulative energy dissipation density and its depth‑based Pearson coefficient curve. A Physical Knowledge Module (PKM) is embedded, encoding deformation growth, stress relaxation, and damage propagation into a composite loss function that combines data loss and physical loss. The framework is applied to the SJLS tunnel project. Bayesian optimization and Dirichlet sampling are employed for hyperparameter tuning. Results show that the model achieves high prediction accuracy (R² &gt; 0.99), outperforming standalone KAN, LSTM, and other baselines. The PKM enhances both predictive accuracy and physical consistency. The proposed framework serves as an efficient tool for real‑time prediction of multiple time‑dependent responses and for mechanical parameter inversion based on multi‑source monitoring data.</p>