Machine learning in quaywall design

Abstract: Today we live in a world where technology is changing the world and projects around us at a rapid pace. It is believed companies will have to change their operations to maintain an edge. At Movares, a Dutch engineering consultancy firm, they recognize the importance of digital transformation. Their goal is to apply digital transformation to their day-to-day operations enabling engineers to focus on innovation. One of these operations concerns the optimization of quay wall designs. In this thesis, an automated optimization routine for the design of qual walls is suggested. The design of Quay walls is influenced by site-specific variables and design variables. Currently at Movares, the design variables are determined based on engineering judgement a combination of norms, experience, and data. The lack of an integral analysis of the design variables makes it difficult to judge the efficiency of the designin terms of costs. Furthermore, the speed of this trial-and-error based approach is limited by the designer interacting with the analysis software. The automated optimization routines suggested in this thesis try to pose a solution to these problems. In an automated routine, the task of the engineer shifts from evaluating results to formulating an optimization problem. The engineer operates at a higher level and an algorithm is responsible for evaluating the intermediary results. The proposed routines can be best described as a databased or data-driven routine and a hybrid routine. The databased routine bases its evaluation solely on data and relies on statistical tools to extract insight. For the design of quay walls, this data includes soil properties, soil geometry and design variables. The hybrid optimization routine combines the use of data with a theory-based model. A theory-based model in contrast to databased models is based on scientific understanding of a system or process, e.g., determining slope stability with Bishop’s method, or soil behavior under cyclic loading. From the work in this thesis, it is shown hybrid optimization routines were able to identify the optimum with respect to costs within an acceptable timeframe. With the use of Machine learning techniques, the total computation time was significantly reduced compared to uninformed sampling techniques.