Machine learning (ML) offers considerable promise for advancing metal processing, such as additive manufacturing (AM). Metal AM can enable the fabrication of components with designed microstructures and hence engineered performance characteristics. By emulating complex correlations within process data, ML allows for the efficient optimization of processing parameters, improved defect prediction, and accelerated real time defect monitoring.

In close collaboration with the IFW group ML for materials science (Dr. Dmitry Chernyavsky, Dr. Denys Kononenko; Institute for Theoretical Solid State Physics), we specifically focus on leveraging ML in the context of laser powder bed fusion, which represents one of the most widely adopted and industrially relevant metal AM technologies. This LPBF process is inherently complex, because it involves interactions between laser parameters, powder characteristics, melt pool dynamics, and solidification behavior. All of those interactions critically impact the evolution of the microstructure, especially involving the formation of defects such as pores and cracks. 

Our research aims to utilize ML for the identification of optimal LPBF processing parameters yielding dense material and for the development of advanced real time monitoring approaches with potential for the implementation into closed-loop control systems. Thereby we focus on two main aspects:

(1) Bayesian optimization (BO) frameworks for modelling process–structure–property relationships: Using BO to efficiently predict optimal LPBF parameters for the successful fabrication of high-density components from industrially relevant, but difficult to process alloys. BO is designed for accurately predicting optimal conditions despite the limited data and for multi-objective scenarios, thereby enabling digital materials design.

(2) ML-based real-time monitoring and defect detection: Developing algorithms that analyze sensor data in real time to detect defects, such as cracks or pores during LPBF fabrication. We focus on ML models, such as Autoencoders, for the development of automated defect monitoring systems based on acoustic emissions.