Machine learning for structural design models from the inverse problem perspective

The introduction of novel technologies has historically led to fundamental changes in structural design practices. Despite the increasing prominence of machine learning techniques, the application for structural design models is sparse in current engineering practice when compared to structural analysis models. This publication-based thesis posits that structural design is in fact an inverse problem and presents a methodology to develop accurate, generalisable, and verifiable machine learned structural design models for continuous beam systems. Unique to this perspective is the interlink between structural analysis, design, and optimisation, providing a fundamental shift in engineering philosophy that is conventionally dominated by the forward-problem oriented field of engineering science.

A comprehensive literature review on the range of domains in which inverse problems have been identified in civil and structural engineering is presented first, covering applications in blast engineering, structural health monitoring, and digital twins. This review highlights the extensive use of machine learning in such domains as opposed to traditional optimisation based inverse solvers, and underlines some of the unique advantages which machine learning models provide to help address the current design challenges in industry.

The thesis subsequently embarks on an in-depth investigation to build a non-iterative machine learned structural design model for continuous beam systems within the three subsequent chapters. The first of these three chapters introduces the novel concept of an influence zone of continuous beam systems, which acts as a generalisable and heuristic estimator of the pertinent local loading information relevant for the design of individual members within such a system. The second chapter takes advantage of the influence zone to develop a generalisable, neural network based structural design model to predict cross-sectional properties of a continuous beam system of arbitrary system size. This chapter explicitly frames the machine learned design model from the inverse problem perspective, ideates an appropriate feature selection for the input parameter space, tests various architectures, and evaluates the accuracy of the model. The last chapter utilises a novel physics-informed neural network to evaluate the performance of such an architecture compared to the previously achieved benchmark and showcases the viability of determining the physical accuracy of predictions during inference. This work concludes by reflecting on the novel contributions achieved by this investigation, and the specific scopes for future work.