Prediction of concrete failure behavior under high impact using machine learning

In the field of protective engineering, the rapid and accurate prediction of concrete failure behavior under high impact loads is of great significance. It can help assess the safety of concrete structures. Several methods including empirical formulas and fully connected networks (FCN) were proposed to achieve rapid prediction, but their accuracy was limited. Numerical simulation achieved high accuracy in prediction, but the predicting speed was limited. Developing a method that balances prediction accuracy and speed is urgent. This paper collected data on concrete failure behavior under projectile impact loads, performing feature selection and normalization. A modified residual network (ResNet) structure is proposed to predict concrete failure behavior under projectile penetration. The network replaces the convolutional layers in the traditional ResNet with fully connected layers, removes the pooling layers, and reduces the number of hidden layers, making the ResNet more suited to accommodate the concrete failure data. Meanwhile, SHapley Additive exPlanations (SHAP) analysis identifies the key penetration feature parameters affecting penetration performance; and a branch network is added to the ResNet to handle these significant penetration feature parameters, enhancing the network's ability to predict concrete failure behavior under high impact loads. Compared with empirical formulas and FCN, the presented model achieves higher prediction accuracy and better generalization ability, while also surpasses numerical simulation methods in prediction speed.