3D CNN-based crack propagation prediction in peridynamic concrete models under freeze-thaw cycles

Abstract

Freeze-thaw damage is a critical factor compromising the durability of concrete, rendering the prediction of crack evolution under freeze–thaw conditions essential for evaluating concrete service life. In this study, the relationship between pore frost heaving force and the frost heaving force state is derived, resulting in a peridynamic formulation for the concrete freeze–thaw problem. A three-dimensional convolutional neural network (3D CNN) prediction model, driven by peridynamic theory, is developed. The influence of porosity and pore frost heaving force on the freeze–thaw performance of concrete is systematically analyzed. Based on the proposed model, crack damage predictions over 20 freeze–thaw cycles are carried out. The results indicate that freeze–thaw-induced damage in concrete increases with higher porosity and greater pore frost heaving force. Furthermore, to enhance the accuracy of the 3D CNN in capturing the underlying physical mechanisms, it is recommended to appropriately reduce the number of pooling layers. The developed prediction model demonstrates excellent long-term prediction capability, achieving an accuracy exceeding 92.3%. Compared with traditional methods, the computational efficiency is improved. This study provides an approach for predicting freeze–thaw damage and the remaining service life of concrete in practical engineering applications.