Machine learning approaches for buildable and sustainable bridges

Kapitel i bok, 2025

For bridge design, the multi-objective challenge of buildability and climate impact, could be supported by data-driven methods and machine learning algorithms. The purpose of this study was to test, compare and evaluate supervised, unsupervised and reinforcement learning. First, set-based design was implemented as a data driven approach for data generation. Second, supervised learning with Neural Networks were evaluated as to prediction of CO2. Third, unsupervised learning was carried out by t-SNE dimensionality reduction to identify data clusters. Fourth, a Q-agent was trained with reinforcement learning to propose bridges. The results show that supervised Neural Networks display low error of the predictions. Furthermore, t-SNE could display data clusters. Finally, the Q-agent training indicated that states and rewards need more attention. It can be concluded that the different machine-learning domains show different potential and challenges for the studied problem.