Reliability-Based Assessment of UHPC–Normal Concrete Interface Shear in Bridges: Comparison of AASHTO Shear Friction, Deterministic, and Probabilistic Machine Learning Models

Abstract Reliable shear transfer across the ultrahigh-performance concrete (UHPC)–normal concrete (NC) interface is critical for composite action in bridge rehabilitation and accelerated construction, including overlays, joints, and closure pours. This study develops a reliability-based comparative framework for UHPC-NC interface shear resistance using a curated literature database of direct shear tests. The AASHTO UHPC shear-friction formulation is evaluated alongside deterministic machine learning (ML) models and a probabilistic ML model based on natural gradient boosting (NGBoost), which learns the conditional distribution of interface shear resistance and estimates mean resistance and dispersion. Model interpretation indicates that fiber-related parameters and their interaction with interface roughness meaningfully influence shear behavior beyond variables explicitly represented in shear-friction formulations. Using consistent load and uncertainty assumptions, all resistance models are embedded within a first-order second-moment (FOSM) reliability framework, and database-specific resistance factors are calibrated for comparative assessment. Professional-factor statistics indicate that the AASHTO formulation is conservative on average but has substantially larger dispersion. In the FOSM-based comparison, this mean conservatism does not translate into higher reliability; by contrast, the ML-based models exhibit lower resistance dispersion and can be calibrated to the target reliability level, with NGBoost providing a more favorable bias–dispersion balance. Additional first-order reliability method and Monte Carlo simulation sensitivity checks for the AASHTO formulation show that explicitly preserving the fitted lognormal professional-factor distribution produces higher reliability estimates relative to the FOSM approximation, highlighting the importance of distributional assumptions and uncertainty propagation in reliability assessment. The calibrated factors are conditional on the database and modeling assumptions and are intended for reliability-informed comparison rather than direct use as design values.