Data-efficient Machine Learning for Geometry Modelling in Cold Spray Additive Manufacturing

Additive manufacturing (AM) is widely recognised as a paradigm shift in the nature of future manufacturing as it has demonstrated the potential to offer the variety of benefits that are difficult to achieve otherwise, including: mass-customisation, great freedom of design, waste minimisation and the ability to fabricate complex shapes. However, the commercial integration of AM technologies is still greatly limited, especially High Production Rate AM (HPRAM) technology in metal AM domains. Such limitation is attributed to the lack of process monitoring and quality assurance measures in metal AM, indicating that quality control is a challenge to overcome. One such quality characteristic is geometry accuracy and consistency due to the AM processes being fundamentally complex with numerous process variables and the limited process modelling capabilities. Recently, data-driven machine learning modelling has attracted increasing attention due to its modelling capability without complete physical AM process insight. The downside of purely data-driven machine learning is the necessity of large volume of data for adequate predictive accuracy. This disadvantage is frequently encountered in industrial AM scenarios with foreseen customised and short-run productions, the high-value nature and process monitoring difficulty. The work presented in this thesis explores data-efficient machine learning for the modelling of macro-scale geometric deposit formation in one of HPRAM processes, Cold Spray AM. Herein, single- and overlapping-track cases are focused to demonstrate the proposed modelling approaches. The significance of this thesis is mainly three-fold: (1) exploring the data-driven machine learning modelling approach beyond its current AM use, (2) proposing data-efficient machine learning approach and compare to mathematical and purely data-driven approaches and (3) leveraging existing AM domain knowledge or previously proposed mathematical models to achieve data-efficiency