Recognizing Mechanical Behaviors of Silt-Based Foamed Concretes from Image of Pores Using Deep-Learning Strategy

Abstract

The pore structure of foamed concrete exhibits an inherent interaction with mechanical behavior. This study developed a novel deep-learning (DL) strategy to predict mechanical behavior of foamed concrete based on the image of pore distribution. This framework includes two parts: a convolutional neural network (CNN) for classifying pore images according to mixture conditions, and a multilayer perceptron (MLP) to reproduce stress–strain relationships under various loading paths. The pore image data set was obtained via scanning electron microscopy (SEM), and stress–strain data were from laboratory experiments. Rigorous training and testing demonstrated that the CNN classifies the mixture condition from pore distribution images with an accuracy of 85%, which is used as the input to the MLP. The MLP successfully captures stress–strain responses under uniaxial compression, uniaxial tension, triaxial shear, and isotropic compression, with strong denoising and generalization ability. The R 2 value can reach 93%, demonstrating successful reproduction of the stress–strain curves. Comparison with mathematical descriptions of the critical state and yield stress further demonstrated the practical applicability of the proposed DL framework. The proposed hybrid CNN-MLP framework can effectively predict the mechanical properties of foamed concrete from pore images.