The structural integrity of precast concrete composite beams heavily relies on the new-to-old concrete interface. Existing numerical models and traditional analytical methods struggle to capture the complex, multi-variable interfacial degradation mechanisms. This study proposes an integrated framework combining high-fidelity 3D nonlinear finite element (FE) simulations with explainable machine learning (XML). A mixed interface constitutive model, seamlessly coupling surface-based cohesive behavior with residual Coulomb friction, was established and validated to accurately replicate the full-range progressive damage and frictional slip. Using an orthogonal experimental design, a 53-sample database was generated to evaluate key design variables, including material strengths and interfacial roughness. An eXtreme Gradient Boosting (XGBoost)-based surrogate model successfully mapped the nonlinear relationships between these features and core flexural indicators, achieving an R² exceeding 0.965. The SHapley Additive exPlanations (SHAP) framework was subsequently introduced to decode the algorithmic black box, providing transparent traceability of feature importance. Results reveal that cast-in-place concrete strength dominates initial flexural stiffness, whereas tensile reinforcement dictates yield and ultimate capacities. Crucially, interfacial roughness governs the post-peak response, enhancing energy dissipation by over 400%. This data-driven strategy validates that rationally matching material strengths with optimal interfacial friction maximizes the comprehensive flexural potential of composite members.