Reinforcement Learning Process Control in Powder Bed Fusion Additive Manufacturing

Powder Bed Fusion (PBF), a prominent metal Additive Manufacturing (AM) technique, is an original technique for producing 3D parts by adding material iteratively, on a layer-by-layer basis. In PBF, a heat source melts metallic powder, layer-by-layer, enabling the creation of potentially complex parts with tailored geometries. However, despite this advantage, PBF remains challenging to analyse and understand across multiple scales due to complex, nonlinear thermal phenomena and interactions. This complexity, along with the lack of control-oriented PBF models, hinders the development of closed-loop control systems. As a result, practitioners still rely on empirical openloop settings rather than feedback, which often leads to suboptimal, inconsistent builds. Existing control attempts, based on control theory, have shown promise when applied to simplified PBF models with fixed control targets. However, as part geometries and PBF settings become more intricate, the required models may be inaccurate or even unknown for some aspects of the process, making traditional control methods challenging to design and implement. In contrast, data-driven techniques, such as Reinforcement Learning (RL), offer a more flexible alternative. RL is trained to derive optimal policies through trial-and-error interactions with the control environment, bypassing model assumptions. This flexibility, however, comes at a cost, since RL faces critical challenges: in RL training stability, marked by unpredictable training behaviour and high variance, and in constraint handling mechanisms, essential in safety-critical tasks. This thesis, presented as a coherent collection of publications, aims to bridge the gap between RL algorithms and their practical use in PBF. Specifically, this thesis focuses on addressing the RL stability and constraint challenges in the context of PBF, enabling reliability and broader adoption. The goal is to establish safe and effective RL control in real-world PBF settings, ultimately unlocking the potential of RL in critical AM applications.