Project A07

Anomaly-driven reinforcement learning for process optimization in additive manufacturing of hybrid materials

Due to the complex relationships between material properties, heat exchange, and process parameters in the Laser Directed Energy Deposition (DED-LB) process, it is challenging to develop accurate and computationally efficient physics-based models for real-time process optimization. Machine learning techniques have the potential to enable computationally efficient real-time optimization. The objective of the project is to develop an anomaly-driven reinforcement learning (RL) model for the dynamic adaptation of process parameters in the DED-LB process. The RL model will be trained to adjust the process parameters to minimize anomalies and thus ensure product quality. Experiments will be conducted to identify suitable parameter ranges, analyze their influence on component properties such as porosity, and collect process data. Through statistical analyses, thresholds for classifying anomalies will be defined and the process data will be annotated. Given the data, an anomaly detection (AD) model will be trained to learn the process states of good quality components as normality. In addition, a surrogate model for simulating the process will be trained with process data from components of varying quality. By interacting with this simulation, the RL model will learn to optimize process parameters by minimizing the anomaly scores of the AD model and thus ensure product quality.

Project leaders
Prof. Dr.-Ing. Jan C. Aurich
Jun. Prof. Dr. Sophie Fellenz
Prof. Dr. Marius Kloft

Doctoral researchers
Maik Schürmann, M.Sc.
Saurabh Varshneya, M.Sc.

Publications

Peer-reviewed articles and books

S. Ghansiyal, L. Yi, M. Klar, J.C. Aurich: Designing Porous Structure with Optimized Topology using Machine Learning. Procedia CIRP 125 (2024): S. 190-195. 10.1016/j.procir.2024.08.033