Fabrication Forecasting of LPBF Processes Through Image Inpainting with In-Situ Monitoring Data

Industrializing metal additive manufacturing for mass production requires a consistent manufacturing process that reliably produces high-quality end products. In order to meet these quality requirements, layer-wise in-situ monitoring data is captured to detect process deviations that potentially lead to product defects. However, this way of process monitoring is limited to a retrospective analysis, where defect development is usually unavoidable. Still, accurate forecasting of part fabrication within the same or subsequent layers would allow a timely corrective adjustment of the control strategy. In our work, we formulate the forecasting of part fabrication as an image inpainting problem, where areas of the part that have not been printed yet, are treated as missing regions within layer-wise in-situ image data. We propose to train generative inpainting models to fill in these missing regions, thus predicting possible outcomes of the printing process. In our experiments, we train a generative neural architecture on layer-wise images of heat signatures that were captured with an optical tomography monitoring system during Laser Powder Bed Fusion (LPBF) processes. By varying machine parameter configurations and part geometry, we evaluate the prediction capabilities of the model. Our results reveal, that our model is capable of accurately predicting realistic outcomes of LPBF processes using in-situ monitoring data with a sufficient level of detail. From that we conclude, that generative models show promising results towards an online defect prediction system, that allows a timely intervention of the current control strategy. With our approach, we lay the foundation of a model-based control framework that may prevent product defects from forming.