Process Control for Direct Ink Writing of Composite Materials Via Reinforcement Learning

Abstract

Material-extrusion-based additive manufacturing is a versatile technology attracting interest across multiple industries in energy, semiconductor, biomedical, and construction. It has several unique merits including material variety, the ability to produce complex and customized geometries, and cost-effectiveness. The feedstock material, commonly referred to as the feedstock ink, is a viscous elastomeric fluid (i.e., a kind of composite material). Extrusion of such feedstock inks is a dynamic process governed by multiple interdependent process parameters. The intrinsic interactions within the process often lead to ink flow heterogeneity, resulting in unpredictable under-extrusion and over-extrusion. This ultimately affects the consistency of ink deposition. Traditional control strategies rely on open-loop, empirically derived methods that are often material-specific and nontransferable. Lacking a generic control strategy poses significant challenges with composite material-based inks. In this work, a data-driven reinforcement learning (RL)-based controller is developed to address the challenge. The stochastic state transition technique is adopted to generate the offline data set for the RL-based controller training and reduce the physical experimentation needs. The RL-based controller adjusts the printing in real-time to compensate for the over-extrusion and under-extrusion when such undesired phenomena occur. Several validation material extrusion (i.e., direct ink writing) experiments demonstrate the effectiveness of the RL-based controller. Besides, the RL-based controller is transferable and can be adapted to new process conditions with limited experiments at new process conditions, showing its generalization ability and broadening the application scenarios of the RL-based controller.