Prediction of complex strain fields in concrete using a deep learning approach

Stress-strain analysis is pivotal in high-performance concrete composite design, which has traditionally relied on experiments and numerical simulations. However, these approaches are inefficient due to the astronomical number of possible combinations. This paper thus proposes a high-efficient surrogate model utilizing generative deep learning (DL) for complex strain field prediction to address this research topic. A dataset containing 2000 samples with randomly distributed geometries and corresponding strain fields is built as ground truth. A DL model is constructed, trained, and validated to capture the intricate geometry-strain relationships within the dataset. The predictions to material properties demonstrate astonishing accuracies of 0.54% MAPE in strain fields and of 0.96 R2 in deformability ranking. Furthermore, the proposed approach offers remarkable adaptability and extensibility to varied geometries and failure modes, supported by real-world tests. This method significantly enhances the efficiency of evaluating concrete physical properties by leveraging the geometry for direct strain field analysis.