Geometry-agnostic Melt-pool Homogenization of Laser Powder Bed Fusion through Reinforcement Learning

This work explores the feasibility of a geometry-agnostic laser power control strategy for laser powder bed fusion (L-PBF) using reinforcement learning. The controller is designed to anticipate and compensate geometry-induced process inhomogeneities, as well respond to in-process uncertainty through feedback control. To train the reinforcement learning controller, first a reduced-order simulation model is fit to experimental data. Then, the optimal control strategy is found through reinforcement learning on this reduced-order model. After the training, we demonstrate that the learned control strategy can reduce up to 55% of the error 2-norm and 59% of the standard deviation with respect to a given reference value. Moreover, the learned control strategy is applicable to novel build geometries without any additional tuning, or modification of the controller, in which we find that the controller attenuated 2-norm error by 62% and variation levels by 60% when deployed on a new (test) geometry, presenting the efficacy of the proposed controller. Finally, the experimental validation of the algorithm in a 'playback' setting resulted in a 24% reduction of both 2-norm error and variation levels, highlighting its potential in an industrial L-PPBF system.