Abstract & Overview
This cumulative thesis presents novel methodologies that harness artificial intelligence (AI), particularly probabilistic and generative deep learning (DL), to support the assessment of existing structures and the structural design of new ones, both with a primary focus on bridges. Bridges are indispensable to transportation infrastructure, yet many in Western countries are approaching or exceeding their intended service life. This creates a pressing need for efficient methods to assess their condition, extend their lifespan, or replace them with new structures. The first part of the thesis addresses the structural assessment of existing bridges. To overcome the limitations of current time- and resource-intensive assessment procedures, a probabilistic DL framework based on Bayesian neural networks is proposed. Trained on high-fidelity non-linear finite element simulations of reinforced concrete frame bridges, the model delivers code compliance predictions efficiently with calibrated uncertainty estimates. The results demonstrate the potential of data-driven surrogates to act as scalable pre-assessment tools, offering decision support for infrastructure maintenance planning. The second part explores how AI can augment the structural design process. Bridge design typically relies on expert intuition and iterative modelling, which can limit workflow efficiency and the optimality of the structural solution. Current software supports only time-intensive forward analyses of predefined designs. In contrast, this work proposes a generative DL framework that enables inverse design: generating valid design alternatives directly from specified performance targets. A proof-of-concept study on a pedestrian bridge project demonstrates this capability through a conditional variational autoencoder (CVAE) that allows both rapid performance prediction and conditional design generation within a defined parameter space. Building on these findings, the AI-eXtended Design (AIXD) toolbox was implemented: a domain-agnostic, low-code Python library that allows domain experts to train and deploy such DL models for any vector-parameterised design task. To demonstrate the benefit of pre-trained DL models that are applicable to a wide range of real-world project scenarios, two case studies are presented: (i) the design of reinforced concrete frame bridges for railway underpasses and (ii) the configuration of a modular bridge system with a steel-concrete composite deck for highway structures. Across these applications, the DL models learn continuous approximations of the complex design spaces and propose diverse sets of structurally viable, material- and cost-efficient configurations while complying with the individual project constraints. In addition, they provide real-time quantitative performance feedback and enable efficient sensitivity analysis to guide design refinements by experts. Overall, this thesis establishes a methodological and computational framework for AI-augmented structural engineering and demonstrates that AI can serve as a powerful analytical and generative tool, positioning AI models as co-pilots that augment human expertise. By combining computational efficiency with human contextual understanding, the proposed framework supports performance-driven engineering decisions and accelerates established workflows, ultimately highlighting the benefits of reporting and harnessing structural databases within the architecture, engineering, and construction sector.