Data-driven lumped plasticity hysteresis model for predicting the cyclic response of rectangular reinforced concrete columns

초록

Accurate prediction of the cyclic response of reinforced concrete (RC) columns is essential for reliable seismic performance assessment. Conventional physics-based models often require extensive calibration and involve con siderable computational cost. This study proposes a data-driven lumped plasticity framework that integrates the modified Bouc-Wen-Baber-Noori (mBWBN) hysteretic model with machine learning to predict the cyclic response of RC columns. The framework is developed and validated using an experimental database of 314 rectangular RC columns, which covers a wide range of geometries, reinforcement layouts, material strengths, and axial load ratios. For each specimen, the measured force-displacement response is fitted to the mBWBN formulation using ant colony optimization to identify specimen-specific hysteretic parameters. Bayesian-tuned XGBoost models are trained to map 12 structural features to these parameters, which allows for rapid, calibration-free prediction of the full cyclic response. Validation results show that the proposed approach reproduces global hysteretic behav ior and captures cyclic stiffness degradation, strength deterioration, pinching effects, and unloading-reloading characteristics with reasonable accuracy, particularly for flexure-controlled columns. The framework also closely approximates peak strength, effective stiffness, and cumulative energy dissipation across the full loading history, with a computational cost reduction of more than 50% compared with fiber-based simulations.

키워드