Machine learning-enabled characterization of concrete mechanical strength through correlation of flexural and torsional resonance frequencies

Abstract

In this study, an assessment of concrete compressive strength was conducted using an impulse excitation data-driven machine learning (ML) framework. The model was constructed based on a deep neural network and aided by the backpropagation method, ensuring a precise training process. In contrast to prior research, which mainly focused on mixture components, a meaningful relationship between physical parameters—resonant frequencies and elastic moduli—and compressive strength was established by our ML model. Remarkable performance was demonstrated, with a root mean square error of 2.8% and a determination coefficient of 0.97. Through Pearson analysis, correlations between input features and output targets, ranging from -0.29 to 0.90, were revealed. Notably, the strongest correlations with compressive strength were found in Young's and shear moduli, derived from flexural and torsional frequencies, highlighting the pivotal role of dynamic elastic response in concrete's mechanical behavior. Furthermore, the findings indicated slight prediction deviations in cases involving samples with a high Poisson's ratio. This work illuminates the potential for accurate compressive strength prediction by leveraging concrete's dynamic response, particularly flexural and torsional modes, thereby opening avenues for research into concrete compressive strength without direct consideration of sample ingredients.