Additive manufacturing (AM) of metal products is achieved by selective laser melting (SLM) at the surface of a metal powder bed. SLM locally fuses the powder to construct a thin object layers (~30μm). The object then sinks in the bed and a wiper recoats the top of the object with an equally thin layer of fresh powder. The laser then adds another layer and the process is repeated until the object is fully formed. To fulfill the need for quality assurance, without tedious post-process quality control, a prediction model using process monitoring data to uncover defects is developed in this use-case. Even though the process is fast, sensing is the easy step for in-process monitoring. The difficulty lies in the development of algorithms to detect, predict, and ultimately prevent defects. Handling massive amounts of data in combination with real-time processing forms a technical challenge, as does the correct interpretation of measured data in predicting the presence of defects and their location. The ambition of CoE RAISE for the AM use case is to upgrade machine learning models to predict defects from shallow to deep models, and to upgrade training data to multiple modalities (sensor fusion) and by at least an order of magnitude in volume. Thereby, the error rate in keyhole porosity prediction is reduced by a factor of two or more and the robustness of the model for different manufacturing parameters is increased. This enables quality certification for porosity count based on build-time measurements only, eliminating a key barrier for increasing AM industrial output of high-value products. The data and model are analyzed to investigate more thoroughly the manufacturing parameter space, and to distill the model into a smaller, real-time capable model. The distilled model supports the real-time control of the process, considerably reducing the defect rate.