Distributed <scp>AutoML</scp> framework for multi‐objective optimization of concrete crack segmentation models

Abstract Monitoring cracks in concrete surfaces is essential for structural safety. While machine vision techniques have received significant interest in this domain, selecting optimal models and tuning hyperparameters remain challenging. This paper proposes a Distributed Automated Machine Learning (AutoML) framework for efficiently designing and optimizing deep learning models for concrete crack segmentation. The framework simultaneously optimizes multiple objectives, including segmentation accuracy and mean intersection over union (mIoU), while minimizing model parameters, depth, and computational complexity. Applied to the optimization of five U‐Net model variations, the framework is compared against non‐dominated sorting genetic algorithm (NSGA‐II), Gene Manipulation Genetic Algorithm (GMGA), grid search, and random search. Results show that the proposed method produces 55% to 177% more optimal models, with 79% to 99% of its models outperforming those generated by alternative approaches. The findings highlight the potential of distributed AutoML for real‐world structural health monitoring applications, especially in edge devices.