Metal powders are used as feedstock in a variety of manufacturing methods, wherein the metallic particulates may be deposited upon a substrate, spread upon a build plate, or melted layer-by-layer. Regardless of how the powder is consumed during a given manufacturing process, the ease with which a powder flows maintains a significant impact on the properties and behavior of the resultant component procured. This is especially true for cold spray additive manufacturing, which is a solid- state metallic particulate deposition process that requires suitable flowability to prevent component waviness as well as prevent nozzle clogging.

That being said, particle-level geometrical properties, such as size and shape features, greatly impact the capacity of a powder to flow, or its flowability. Existing techniques for measuring a particle’s flowability require manual powder processing which can be both time-consuming and vulnerable to man-made error. Both of these setbacks can be addressed through automatic, data- driven techniques trained to classify flowability from particle property measurements and expertly procured powder Hall Flow rates for training and testing.

To our knowledge, we present the first computational model for classifying flowability (when captured in terms of the industrially standardized Hall Flow rates) of a metal powder given its particle-level physical property measurements. After obtaining twenty-one metal powder samples from manufacturers, particle- level physical property measurements of each powder were measured. We used these particle measurements to train a Decision Tree model to classify whether a particle’s Hall Flow rate was fast or slow. With an achieved accuracy 98.04%, our work illustrates that powder flowability, through the lens of Hall Flow rates, can be accurately classified using data-driven techniques.

This work lays a promising foundation for further exploration into predictive models for this cutting-edge domain, this technique only views the prediction of powders from the particle level. In our dataset, each powder is composed of multiple particles, each with their own set of features. Current machine learning algorithms do not accurately model the relationship between multiple-instance data and their labels. Thus, we built a series of Multiple-Instance Regression frameworks to process multiple-instance powder data so that it can be fed into common machine learning models, such as a Random Forest classifier. In addition, we also developed a novel strategy for augmenting multiple-instance data to offset the lack of resources and data available in the Materials Science field and the lack of augmentation techniques for multiple-instance data.