Gaussian process regression-based load-carrying capacity models of corroded prestressed concrete bridge girders for fast-screening and reliability-based evaluation

Abstract

Fast screening and reliability-based evaluation of corroded prestressed concrete (PC) bridge girders for safety assessment requires efficient capacity prediction models. This paper aims at developing data-based prediction models to quantify corrosion effects on load-carrying capacity of PC girders that are precast and standardized for short-span bridges. The load-carrying capacity of corroded bridge girders depends on their initial design configuration (e.g., those characterized by the standardized PC girders), loading condition (e.g., shear-span-to-effective-depth ratio), and corrosion conditions (e.g., corrosion degree). To develop data-based prediction models, generating sufficient data through physical experimental tests of full-scale corroded PC girders is prohibitively costly. Thus, a virtual experimental database is generated numerically using two-dimensional continuum-based finite element (FE) models for corroded PC girders, after being validated using 9 corroded PC girders tested in the literature with flexure or shear failure. To this end, a total of 4,165 PC girders under point loading are simulated to consider various design, loading, and corrosion conditions to estimate their load-carrying capacities. With this database, probabilistic machine learning, specifically Gaussian process regression or kriging, is used to develop (1) a residual capacity factor model (i.e., the ratio between load-carrying capacities from corroded and uncorroded girders), and (2) a load-carrying capacity prediction model (i.e., to predict the load-carrying capacity of uncorroded/corroded girders). The first model is used to study the load-carrying capacity reduction of PC girders due to corrosion compared to their uncorroded counterparts; this can reveal how the increasing corrosion condition affects the load-carrying capacity of corroded PC girders together with other design and loading parameters. The second model is developed to predict the load-carrying capacity of corroded PC girders as a function of design, loading, and corrosion parameters; its application to conditional reliability analysis provides insights into corrosion effect on the probability of failure of corroded PC girders at given load levels when the capacity is affected by corrosion. The application results of the two models enable engineers to quantify the corrosion effect on PC girders in terms of (1) reduction in load-carrying capacity for fast screening and (2) increase in probability of failure for reliability-based evaluation.