Enabling the characterization of powder and part through the Powder Bed Fusion Additive Manufacturing Pipeline

 The Additive Manufacturing (AM) pipeline is a rich source of data, going from the powder feedstock to the final part. Powder properties such as size and morphology dictate powder behavior in the AM machine, which when combined with machine variables such as processing parameters, heavily influence the part properties like surface roughness and part density. With such a rich pipeline and many variables to contend with, it is a challenging task to draw generalized conclusions about the performance of powders and the influence of parameters on part outcomes. Added to this challenge is the lack of standardization in testing and experimentation with diverse sets of parameters. This can come from an insufficient understanding, for example, of powder behavior under AM process conditions and the physical properties that drive it. It is also due to the vast number of variables that can be manipulated, which causes a bottleneck in terms of experimentation time. Metal powders are extensively used in additive manufacturing processes which necessitates standardized characterization methods for powder properties. Understanding both powder flowability as well as morphology and the interconnected relationships between the two is crucial for the choice of powder and in the design of the powder handling systems. The use of computer vision and machine learning tools in the additive manufacturing domain has enabled the quantitative investigation of qualitative factors like powder morphology, which affect the flowability in AM processes. In this work, flowability is measured through rheological experiments conducted with the FT4 rheometer, through the rotating drum experiments with the GranuDrum, as well as the Hall/Carney flowmeter. The resulting metrics are dissected and their applicability to AM scenarios is discussed. The correlations between these flow metrics and the commonly used particle size and distribution metrics are also explored. While this is necessary for the basic interpretation of flow behaviors, there is a need to delve deeper into the morphological features of powders. Convolutional Neural Networks (CNNs) are used to generate feature descriptors of the powder feedstock, from SEM images, that describe not just the particle size distribution but also the sphericity, surface defects, and other morphological features of the particles. They are then clustered according to morphological similarity, to obtain a powder morphology distribution (PMD) for each powder system. The powder morphology distributions are linked back to their size descriptors for interpretation as well as the flow metrics to understand the effects of powder properties on powder behavior. This analysis approach is intended to be agnostic to the type of AM process and can be adapted to various powder-forming techniques. With respect to the relationships between processing parameters and part properties, a challenging problem related to the optimization of surface roughness is also discussed. While there has been some light shed on how surface roughness drives part properties, it is not quite clear how the processing parameters influence surface roughness. The relatively easier task of characterizing vertical surface roughness and the influence of the processing parameters on this type of rough surfaces is discussed. An example of using CNNs to characterize surface roughness and distinguish between differently oriented rough surfaces is presented. In this manner, this thesis sets up frameworks to implement various computer vision and machine learning tools at various parts of the AM pipeline.