Track: Doctoral Dissertation Competition

Abstract

Modern manufacturing technologies such as Additive Manufacturing are a vital pillar of Industry 4.0 and Smart Manufacturing systems. One of the most promising AM techniques is Directed Energy Deposition (DED) which uses a thermal source to generate a melt pool on a substrate into which metal powder is injected. Laser Cladding (LC), a DED additive process, is widely used in machine parts repair and functional coating due to its advantages, such as the low heat input into the substrate. Compared to conventional welding, this results in less induced stress, lower dilution rate, small heat-affected zone, and good metallurgical bonding between the coating and the substrate. The LC technology is not as robust or standardized. The part quality, mechanical properties, and microstructure of Laser-Cladded parts are neither as commercially predictable nor controllable. This research aims to develop Deep Learning models that predict and classify the LC process for LC coating and single and interfering-bead samples. The ML models will detect faulty process input parameters and predict the quality. This detection and prediction take place in real time, which is vital in producing the additive process’s data model component of a Digital Twin (DT). Digital imaging of a coaxial CMOS camera and melted pool temperature measurements using a LASCON LPC04 pyrometer are being used as in-situ sensory devices.

A plethora of articles have limited their ML models to individual combinations of substrate and powder and focused on the three main process parameters (laser power, laser beam scanning speed, and Powder feeding rate). This research project expands the scope to include a different substrate and powder combinations, utilize additional parameters, and exploit different powder/substrate combinations. The Inputs will be gathered and manipulated in a controlled environment to predict both good-quality and poor-quality LC. The process will take place by building geometries selected strategically to associate uncontrollable parameters, especially the temperature of the substrate, in the prediction model. The online optical images and the offline microscopic analysis of the sectioned bead are mapped together. In a later stage, the ML can potentially integrate into a control system and digital twin models.

Keywords

Additive Manufacturing; Laser Cladding; Prediction; Machine Learning; Melt Pool; Digital Twin.