Interpretable Data-Driven Explicit Shear Formula for Keyed Dry Joints Considering Steel Fiber Contribution

Precast concrete segmental bridges rely on keyed dry joints to transfer shear forces, yet existing formulas have insufficient accuracy for normal concrete (NC) and ultra-high-performance concrete (UHPC) joints. In this study, a unified database containing 158 specimens of keyed dry joints was established and systematically screened, and 134 specimens were retained for machine-learning modeling and formula development. Among six machine-learning models, Categorical Boosting (CatBoost) showed the best prediction performance. SHapley Additive exPlanations (SHAP) analysis indicated that confining stress, concrete compressive strength, the area of the base of all keys in the failure plane, and the area of contact between smooth surfaces on the failure plane govern shear capacity. Based on these key variables, multiplicative and additive explicit formulas were developed, and the additive formula showed better overall performance, with average experimental-to-predicted ratio (Avg), coefficient of variation (CoV), and R2 values of 1.08, 0.22, and 0.94, respectively. Although the additive formula stably predicted NC specimens, it showed insufficient accuracy for UHPC specimens. After introducing steel fiber volume fraction into the root shear term, the UHPC subset’s Avg and CoV improved from 1.30 and 0.24 to 1.09 and 0.21, respectively. The proposed framework integrates data-driven prediction and physical interpretability, providing an engineering-oriented method for shear-capacity prediction of keyed dry joints.