Concrete’s durability is fundamental to the longevity and safety of infrastructure, yet traditional assessment methods are often destructive and time-consuming, necessitating the development of efficient, non-destructive predictive models.This research introduces a novel machine learning approach that leverages high-resolution scanning electron microscopy (SEM) images to predict the durability of concrete by analyzing its microstructural characteristics. A data set of 10,000 SEM images was collected, and a tailored convolutional neural network (CNN) model was used,achieving an accuracy of 94.2% on the test set. The CNN served as a feature extractor for a random forest classifier, leading to significant performance improvements, evidenced by a Cohen’s Kappa coefficient of 0.92, indicating excellent agreement with actual durability classes. The model identified key micro-structural features that influence durability, with pore size distribution being the most crucial, receiving an importance score of 0.376. Additionally, the machine learning-based model out-performed traditional methods such as the Rapid Chloride Permeability Test (RCPT), which had an accuracy of 78.6%, providing a 15.3% accuracy improvement and reducing prediction time from hours to mere seconds (2.5 seconds). This research highlights the potential of advanced imaging and machine learning techniques to enhance concrete durability assessments, facilitating faster, more reliable predictions that can significantly benefit construction practices and infrastructure sustainability.The successful development of a prototype real-time durability monitoring system, with an average accuracy of 92.6% in field trials, underscores the practical applicability of the proposed model in real-world settings.