Zoning additive manufacturing process histories using unsupervised machine learning

Metal additive manufacturing (AM) is currently a highly active area of research within the materials processing and manufacturing community, owing to the promises of lower lead time, increased design flexibility, and location-specific process control. These benefits are balanced by a typically complex processing space, resulting in difficulties when attempting to generalize process-structure relationships across different component geometries. We develop a procedure for reducing the overall complexity of a representation of the AM processing space using techniques from time series analysis and dimensionality reduction. This procedure is highly generic and applicable to a variety of time sequenced signals. We then utilize this reduced feature space as input to a cluster analysis, producing zoned maps of process history. We apply the approach to several canonical geometries, describe its overall utility, and discuss the implications of the resulting zonings in terms of the underlying digital process.