Machine learning prediction of recoater damage topography deviations

In-situ monitoring in laser powder bed fusion (PBF-LB) presents a paradigm for progress towards born qualified parts. This technology has proven useful in many applications such as monitoring for geometric error, layer-wise part defects, and spreading defects. The significance of spreading defects is particularly understudied, especially in the experimental domain. Recoater damage can be particularly detrimental to mechanical performance, as it lends to topography deviations in the build, which could cause porosity, geometric inaccuracies, and potential build failure. Yet, prior literature has not addressed machine learning's ability to predict the severity of recoater damage. This work used multiple feature-based and image-based machine learning algorithms combined with in-situ layer-wise monitoring to predict the amount of topography deviation within recoater damaged sections. The height and width of the topography deviations were measured after the spread profile was exposed to multiple different sizes of recoater wear at different recoater spread speeds and layer thicknesses. The acquired images had different image filtering methods applied to see if a particular image filtering method can increase prediction performance. Ultimately, the image-based machine learning methods showed the best performance when combined with noising filters. In all, this work seeks to find the ideal configuration for the prediction of topography height and width deviations when the powder bed is exposed to recoater damage.