TY  - EJOU
AU  - Guo, Fengqi 
AU  - Tan, Song 
AU  - Jiang, Liqiang 
AU  - Guo, Wei 
AU  - Jiang, Lizhong 

TI  - Interpretable Machine-Learning Prediction of Seismic Performance of Earthquake-Damaged CFRP-Repaired Hollow Bridge Piers
T2  - Computer Modeling in Engineering \& Sciences

PY  - 2026
VL  - 148
IS  - 2
SN  - 1526-1506

AB  - Rapid post-earthquake recovery of high-speed railway bridges requires practical tools for evaluating repaired seismic performance across multiple indicators, rather than peak strength alone. This study develops a target-wise machine-learning framework for rapid multi-indicator prediction of the seismic performance of earthquake-damaged hollow bridge piers repaired with carbon fiber-reinforced polymer (CFRP). An OpenSeesPy-based finite-element model was first validated against test results previously reported by the authors and then used to generate a numerical database covering different loading directions, pre-repair damage states and CFRP repair configurations. Four performance indicators were extracted from the simulated cyclic responses: peak lateral strength, equivalent viscous damping ratio, re-centering indicator and stiffness degradation indicator. Seven nonlinear regression algorithms were compared separately for each performance target using a predefined five-fold cross-validation protocol. The selected estimators were retained in a common four-model predictor bank comprising independently trained target-specific models. The results show that CatBoost was selected for peak lateral strength and achieved an independent-test coefficient of determination (<i>R</i><sup>2</sup>) of 0.9981. Support vector regression (SVR) with a radial basis function (RBF) kernel was selected for the equivalent viscous damping ratio, re-centering indicator and stiffness degradation indicator, with <i>R</i><sup>2</sup> values of 0.9666, 0.9548 and 0.9625, respectively. Several alternative models nevertheless achieved closely comparable accuracy for the latter three indicators. Grouped error analysis showed low percentage errors across the damage-state and loading-direction subsets, although prediction performance varied among several direction–indicator combinations. Model-agnostic interpretability analysis further revealed distinct target-specific attribution patterns for damage state, loading direction, and CFRP parameters across the four performance indicators. These findings demonstrate that target-specific surrogate modelling can capture the coupled but non-identical responses governing repaired seismic performance and can support rapid preliminary screening of CFRP repair schemes within the sampled numerical response domain.
KW  - Earthquake-damaged CFRP repair; hollow bridge piers; machine learning; finite element simulation

DO  - 10.32604/cmes.2026.086426