Machine learning–assisted drift capacity prediction models for reinforced concrete columns with shape memory alloy bars

Abstract

Despite notable progress made in predicting the drift capacity of reinforced columns with steel bars, these techniques and methods are proven inapplicable for accurately predicting the drift capacity of RC columns reinforced with shape memory alloy (SMA) bars. This study employed machine learning (ML) to predict and design the drift limit state of concrete columns using SMA bars. To this end, a total of 292,000 synthetic data points were generated through numerical simulations given the limited amount of experimental data for SMA bars. The data analysis results suggest that the light gradient boosting (LGB) algorithm achieves the best performance in terms of computational efficiency and prediction accuracy among nine candidate ML algorithms considered in this study. Further refinements for the LGB algorithm is introduced to yield better prediction results: (1) Hyperparameters are tuned using particle swarm optimization with an improved particle updating strategy and (2) the dimensions of the input data are reduced using a modified recursive feature elimination algorithm with memorizing capability. In addition, this study demonstrated the application of the proposed ML-assisted drift capacity prediction model to the design of SMA-reinforced concrete columns using modified particle swarm optimization that can help structural designers worldwide.