Sequential, Batch and Multi-Objective Bayesian Design of Experiments for Additive Manufacturing Processes

Research and development of new manufacturing processes require large investments of time, labour, and resources to develop new products. Traditionally, experimentation for manufacturing has been handled using Design of Experiments or Experimental Design (DoE) such as: OVAT, Factorial Designs, and Taguchi Orthogonal Arrays. Whilst these DoE approaches are simple to implement, they select their full experimental budget prior to experimentation which can lead to excessive costs in areas of low experimental value. In new manufacturing industries such as Additive Manufacturing (AM) experimentation is becoming increasingly expensive, whereby traditional DoE approaches will lead to increased costs. The research challenge is to develop data-efficient DoE methods that iteratively select experiments to maximise information gained whilst minimising the number of experiments to improve the understanding of the underlying processes and/or locate the global optimum. This thesis investigates the use of Bayesian Optimisation (BO) as a data efficient DoE method for Expensive AM DoE problems and subsequent development of two novel algorithms, Batch Bayesian Experimental Design Optimisation (BB-DoE) and Multi-Objective Batch Bayesian Experimental Design Optimisation (MOBB-DoE). They are then assessed against current state-of-the-art algorithms and/or applied onto expensive AM DoE case studies. To assess the methodological viability of BO an exploratory investigation is presented. The investigation utilised synthetic benchmarks, theoretical property analysis, and an AM case study to demonstrate BOs capabilities. This investigation also produces the necessary information to contrast each BO cost function through analysis of their properties and performance to determine the suitable cost function to act as the foundation for further algorithmic development. Both BB-DoE and MOBB-DoE algorithms were shown to have improved or comparable performances against current state-of-the-art algorithms on synthetic problems. BB-DoE also demonstrated favourable performance on an AM case study for a Directed Energy Deposition (DED) process seeking an optimal Dendritic Arm Spacing (DAS) property by locating the optimal DED settings.