Resumen de Optimal deep learning assisted design of socially and environmentally efficient steel concrete composite bridges under constrained budgets
Infrastructure design is strongly influenced by the search for solutions considering the impact on the economy, the environment, and society. These criteria were strongly related to the definition of sustainability by the Brundtland Commission in 1987. This milestone posed a challenge for technicians, scientists, and legislators alike. This challenge consisted of generating methods, criteria, tools, and regulations that would allow the inclusion of the concept of sustainability in developing and designing new infrastructures. Since then, small advances have been made in the search for sustainability, but they need more in the short term. As an action plan, the United Nations established the Sustainable Development Goals, setting the year 2030 as the target for achieving them. Within these goals, infrastructure is postulated as a critical point. Traditionally, methods have been developed to obtain optimal designs from the point of view of economic impact. However, although recent advances have been made in implementing and using complete life cycle analysis methods, there still needs to be a clear consensus, especially in the social pillar of sustainability. Given that sustainability encompasses different criteria, which in principle do not necessarily go hand in hand, the problem of finding sustainability is posed not only as an optimization problem but also as a multi-criteria decision-making problem.
The main objective of this doctoral thesis is to propose different methodologies for obtaining optimal designs that introduce the pillars of sustainability in the design of steel-concrete composite bridges. A three-span box-girder bridge is proposed as a representative structural problem. Given the complexity of the structure, which involves 34 discrete variables, optimization with mathematical methods is unaffordable. Therefore, the use of metaheuristic algorithms is proposed. This complexity also translates into a high computational cost for the model, so a deep neural networks model is implemented to allow the validation of the design without the need for computation. Given the problem's discrete nature, discretization techniques are proposed to adapt the algorithms to the structural optimization problem. In addition, to improve the solutions obtained from these discrete algorithms, hybridization methods based on the K-means technique and mutation operators are introduced depending on the type of algorithm. The algorithms used are classified into two branches. The first are those based on trajectories such as Simulated Annealing, Threshold Accepting, and Old Bachelor Acceptance. Moreover, swarm intelligence algorithms such as Jaya, Sine Cosine Algorithm, and Cuckoo Search are used. The Life Cycle Assessment methodology defined in the ISO 14040 standard is used to evaluate the social and environmental impact of the proposed designs. The application of this methodology allows the evaluation of the impact and comparison with other designs. The single-objective evaluation of the different criteria leads to the conclusion that cost optimization is associated with a reduction of the environmental and social impact of the structure. However, optimizing environmental and social criteria does not necessarily reduce costs. Therefore, to perform a multi-objective optimization and find a compromise solution, a technique based on Game Theory is implemented, proposing a cooperative game strategy. The multi-criteria technique used is the Entropy Theory to assign criteria weights for the aggregate objective function. The criteria considered are the three pillars of sustainability and the constructive ease of the top slab. Applying this technique results in an optimal design concerning the three pillars of sustainability and from which the constructive ease is improved.
