Seismic performance evaluation of continuous reinforced concrete bridge structures based on the integrated approach of machine learning and symbolic regression

Abstract

In recent years, the performance-based seismic assessment method using general decision-making indicators has become the mainstream approach for evaluating the seismic safety of structures. However, there are still some deficiencies when assessing the seismic performance of continuous reinforced concrete bridge structures: 1) Insufficient consideration of the randomness of structural materials and geometric dimensions in stochastic finite element analysis; 2) Subjectivity in selecting and constructing seismic intensity measures (IM) during component fragility analysis, with significant variability in the functional relationship between IM and Engineering Demand Parameters (EDP); 3) Lack of a database for repair measures and decision criteria (such as repair time and repair cost) for bridge component categories and damage states based on national specifications. The study aims to investigate the seismic performance of a continuous beam bridge by considering the randomness of material and component parameters, as well as the randomness of seismic ground motion, through stochastic finite element analysis. Machine learning and symbolic regression algorithms will be employed to perform component-level fragility analysis for different bridge elements. Targeting a continuous girder bridge, this study conducts stochastic finite element analysis incorporating material/component parameter variability and ground motion randomness. Machine learning algorithms (Ridge Regression, K-means/CLA clustering, Pearson/Spearman/Kendall correlation coefficients, Distance/Maximal Information Coefficients) combined with symbolic regression via Genetic Programming (GP) are employed for component-level fragility analysis. A repair strategy database incorporating cost and time metrics is subsequently established to assess bridge seismic performance using decision indicators (repair/reconstruction costs, downtime). The results indicate that bridge repair costs and time exhibit pronounced nonlinear growth with increasing PGA. At low intensities (PGA < 0.5 g), longitudinal-direction costs slightly exceed transverse-direction values (four components, e.g., piers, contributing 90 % of total costs), while transverse-direction repair complexity surges at high intensities (PGA > 0.5 g) due to shear key failures. The coefficient of variation (COV) peaks at moderate-high intensities (0.5–0.8 g) and declines thereafter as structural collapse modes converge. Critical cost-effectiveness equilibrium thresholds are identified at 0.92 g (transverse) and 0.93 g (longitudinal), beyond which repair costs (>1.5171 million yuan) surpass reconstruction expenses, serving as pivotal criteria for post-earthquake decision-making.