Identification of experienced temperature in mortar and concrete using microstructural image and deep learning

Abstract

This study presents a framework for identifying the temperature experienced by fire-damaged mortar and concrete using scanning electron microscopy (SEM) images and deep learning. A dataset of 16,484 SEM images of cement paste mixes with varying water-to-binder ratios and pozzolanic materials, exposed to temperatures ranging from 200 to 800 °C was established. Then a deep-learning model based on a convolutional neural network (CNN) for SEM image classification was trained, achieving a high accuracy above 98 %. To test the method's generalizability, cement paste mixture with a different water-to-binder ratio, mortar mixture with sand inclusion, and concrete mixture with coarse aggregates were prepared and exposed to different temperatures. The predicted temperatures deviated from the target temperatures within 8.6 %. Finally, visualization of the deep learning model was used to identify the critical features that influenced the prediction. The outer hydration products with smaller pores had a higher influence on samples before heating, whereas porous dehydrated products were more influential in samples exposed to high temperatures.