Additive manufacturing process parameter design for variable component geometries using reinforcement learning

Ensuring consistent high quality across diverse components in additive manufacturing (AM) necessitates a rigorous and resource -intensive process of trial -and -error experimentation. In practical terms, this entails a substantial investment of time and resources. Addressing this challenge involves the integration of physics -based process simulations with general-purpose optimization algorithms, facilitating proactive process optimization. This strategy effectively directs costly experimental endeavors toward the most promising variations. However, a significant limitation of this approach is the substantial computational time requirement, particularly in the context of iterative optimization. To circumvent the computational constraints inherent in the optimization process, surrogate -based optimization methodologies are commonly employed. These surrogate models are typically custom-tailored to specific scenarios, lacking the capacity to adapt to a diverse range of manufacturing contexts. Consequently, even minor modifications, such as alterations in component geometry, render these surrogate models obsolete, necessitating the labor-intensive processes of data resampling and surrogate model retraining. One highly promising avenue for addressing these challenges involves the application of Reinforcement Learning (RL), a computational technique that seeks to determine optimal actions within dynamic and variable contexts. Within the framework of this research, RL is leveraged to estimate optimal process parameters (referred to as "actions") across a spectrum of component geometries (referred to as "situations"). After the training phase, the model demonstrates a remarkable capacity to furnish meaningful parameter estimations, even when confronted with novel geometries that were not part of the original training dataset. Consequently, it encapsulates transferable insights derived from generic process samples, successfully applying them to the characterization of new and non -generic components. The intrinsic advantage of this approach lies in its ability to harness and recycle extant data, obviating the need for repetitive data collection and model reconfiguration. This pioneering method thus holds profound promise for streamlining both the design of components and manufacturing processes in parallel, ultimately contributing to the enhancement of efficiency and cost-effectiveness within additive manufacturing. Although this study has focused on examining geometries that could be representative of components found in complex industrial settings, in the future more intricate geometries should be considered in the dataset for broader generalizability.