Machine learning techniques for predicting the peak response of reinforced concrete beam subjected to impact loading

Abstract

To meet the growing need for resilient structures in seismic and high-impact zones, accurate prediction of the response of reinforced concrete (RC) beams under impact loads is essential. Traditional methods, such as experimental testing and high fidelity finite element models, are often time consuming and resource intensive. To address these challenges, this study investigates various ensemble and non-ensemble machine learning techniques—including support vector machine, gaussian process regression (GPR), k-nearest neighbor (KNN), gene expression programming, random forest, decision tree, boosted tree, adaptive boosting tree, gradient boosting algorithm, stochastic gradient descent, and artificial neural network—for predicting the peak response of RC beams under impact loads. A set of 145 experimental data points from 12 different sources is used to train and evaluate these machine learning models. Key parameters in the data include beam width and depth, span, reinforcement ratios, concrete strength, steel yield strength, deflection, and impact characteristics. Except for KNN, all models showed satisfactory generalization capabilities with R2 values over 0.8. Statistical errors such as RMSE, a-10 index, MAE, and a-20 index are within acceptable limits. The GPR model is the most effective with R2 value of 0.95. Moreover, Shapely analysis identified beam depth, impact velocity, and beam breadth as critical factors. Overall, this study demonstrates the efficacy of machine learning in accurately predicting the behavior of RC structures under impact loads, providing valuable tools for civil engineers in design and analysis.

Keywords

• Machine learning
• Peak response
• Gene expression programming
• Shapley