Automatic Powder Bed Leveling for Laser Powder Bed Fusion Based on Machine Learning

Abstract Powder bed leveling is a crucial pre-processing step in Laser Powder Bed Fusion (L-PBF) that directly affects the quality of printed parts. Traditional leveling methods rely on manual adjustments based on operator expertise, which are time-consuming, inconsistent, and prone to errors, often resulting in print defects. This study proposes a machine learning-based automated powder bed leveling system to address these challenges. A dataset of 346 powder bed images, collected from 26 leveling experiments, was annotated into eight distinct classes representing various bed adjustments. Advanced preprocessing techniques, including random oversampling, normalization, and cropping, were applied to prepare the dataset. The proposed system utilizes an ensemble of two MobileNetV2-based architectures, enhanced by transfer learning, L2 regularization, and dropout techniques. The ensemble model demonstrated superior performance, achieving a test accuracy of 91.36%, significantly outperforming baseline and refined models. These results highlight the system’s capability to handle class imbalance and complex image classification tasks, paving the way for more consistent and reliable powder bed leveling in Laser Powder Bed Fusion (L-PBF). Future work will focus on expanding the dataset, exploring advanced ensemble strategies, and evaluating alternative architectures to further improve performance.