Predicting peak tensile stress in mesoscale concrete considering size effects: A data-physical hybrid-driven approach

Abstract

The size effect and strength discreteness observed in concrete stem primarily from its mesoscopic heterogeneity. Despite this understanding, establishing a clear relationship between the macroscopic and mesoscopic mechanical behaviors remains a considerable challenge. While some advancements have been made in studying the size effect of concrete, particularly regarding the rate effect, there has been relatively less emphasis on addressing the influence of random meso-component distribution. To investigate the tensile behaviors of concrete across varying low strain rates (10−5 s−1∼10−1 s−1) and model sizes, a meso-model dataset was established comprising double edge notched concrete with mortar matrix, coarse aggregate, and interface transition zone. A convolutional neural network introducing physical parameters was implemented to capture the potential non-linear relationship of concrete from meso-structure, strain rate and model size to tensile peak stress. Subsequently, the data-driven solution for the “Static and Dynamic unified” Size Effect Law (SD-SEL) in concrete tensile strength was developed by deconstructing the back propagation (BP) neural network to enhance its rationality, followed by the verification of the proposed formula. The findings indicate that the Data-Physical Hybrid-Driven approach effectively analyzes the mechanical response of concrete without the need for complex mechanical derivations, offering a promising tool for studying the size effect of composite materials.