Multi-objective process parameter optimization using a hybrid reinforcement/machine learning model in metal additive manufacturing

Additive manufacturing enables the production of complex geometries and lightweight structures. However, producing defect-free components fundamentally depends on the careful optimization of process parameters that influence critical factors, such as melt pool depth, in laser powder bed fusion (L-PBF), an important class of metal additive manufacturing methods. Hence, for the first time, this paper presents a novel data-driven framework that integrates machine learning (ML) with reinforcement learning (RL) to optimize process parameters, ensuring the desired melt pool characteristics while meeting multi-objective design criteria. Experimental data from the literature were extracted to train several ML models, with Gaussian process regression (GPR) featuring a Matérn kernel proved to be the most accurate predictor. This model served as the digital twin of the L-PBF process and was integrated with a Q-learning-based RL agent to derive the optimal process parameters. The GPR model achieved a high prediction accuracy (root mean squared error (RMSE): 34 µm, the coefficient of determination R2: 96.05%) for L-PBF Ti6Al4V alloy using minimal training data. Integrating a Q-learning framework with ML demonstrated superior performance in deriving process parameters that achieved the target melt pool depth while minimizing the normalized enthalpy, directly impacting the formation of the keyhole defect.