Machine Learning-Based Predictions for the Peak Impact Forces of Reinforced Concrete Beams from a Drop Weight

Accurate prediction of the peak impact force of reinforced concrete (RC) structures subjected to impact loads is critical for its impact-resistant design and structural performance assessment. Traditional approaches, such as experimental testing and refined finite element (FE) analysis, are generally time-consuming and resource-intensive. To address these limitations, this study develops an application-oriented machine learning (ML) framework for predicting the peak impact force of RC beams during the impact process. A comprehensive dataset comprising 144 samples is constructed from publicly available drop-hammer impact tests on rectangular simply supported RC beams based on the predefined selection criteria. Five ML models, i.e., Support Vector Regression (SVR), Random Forest Regression (RFR), Gaussian Process Regression (GPR), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANN), are employed to predict peak impact force, and the corresponding model performances are evaluated by comparing with the experimental results. The SHapley Additive exPlanations (SHAP) algorithm and Local Interpretable Model-Agnostic Explanations (LIME) are utilized to assess the relative importance of input features. Based on feature importance, Monte Carlo simulations are further conducted to examine the influence of varying numbers of input features on prediction performance, and a feature reduction strategy is proposed and evaluated. The reliability and effectiveness of the proposed feature reduction strategy are evaluated using five independent experimental cases that are completely excluded from model development, and eleven numerical simulation cases are employed for further supplementary assessment. It is found that: (i) XGBoost achieves the best predictive performance for peak impact force, with a coefficient of determination (R2) of 0.976 on the testing set; (ii) impact velocity and yield strength of longitudinal reinforcement are the most influential features for predicting peak impact force; (iii) the feature reduction strategy with seven input features provides an optimal balance between the model complexity and predictive accuracy for rapid structural design and performance assessment.