Bridging simulation and practice in additive manufacturing: Reinforcement learning for electron beam melting control

Reinforcement Learning (RL) continues to show great promise for intricate process control applications, particularly in areas in which the application model is too complex to design and conventional control methods are challenging to implement. Powder Bed Fusion (PBF) processes, such as Electron Beam Melting (EBM), are prime examples. In such processes, RL offers a data-driven, flexible alternative, whereby process data can be used directly for process modelling and control. However, RL’s practical implementation remains limited by a lack of real-world based testing environments. To bridge this gap, our work includes the development of computationally efficient, real-world based simulation models of the EBM process. These models capture essential thermal dynamics for both simple cuboids and overhang structures while maintaining low computational cost, making them ideal for iterative RL training and evaluation. Using these models, we implement and assess RL control strategies across multiple layers, highlighting RL’s capability to manage critical challenges such as meltpool size consistency. Importantly, this work includes the experimental step prior to RL deployment on a physical machine, ensuring that the RL strategies tested are both effective and safe under conditions that closely mimic the real operational environment. By incorporating considerations of satisfactory performance and RL constraint satisfaction, we demonstrate that RL can achieve reliable performance even in safety-critical EBM settings. Our results underscore the importance of realistic simulation platforms for advancing RL in industrial control, and open a direct pathway towards its adoption in real-world manufacturing systems.