Autonomous learning of digital twins for intelligent extrusion optimisation in MEX

The availability of a digital copy of themselves (digital twin) to experiment upon, may be a fundamental enabler of more autonomous decision-making capabilities for the next generation of intelligent manufacturing machines. A fundamental challenge remains, in that humans are still required to develop digital twins in the first place. Towards a more autonomous, data-driven learning of surrogate digital models, in this paper a new approach is presented where a material extrusion (MEX) machine learns its own extrusion dynamics by autonomously collecting sensor data during trial deposition runs and then develops its own digital twin by automated application of methods of system identification. Through constrained linear least-squares optimisation, the machine is also able to leverage the predictions of the digital twin to optimise its own part program ahead of execution, ultimately producing more uniform, deposited strands. To demonstrate the approach, an experimental campaign is illustrated in which a prototype MEX machine is first instructed to develop its own digital twin of deposition dynamics using data from test deposition runs. After automated optimisation of the part program, performed using the predictor to evaluate improvements, width measurements on subsequent depositions show 35.8 % decrease of width variation (measured as root mean square cumulative error vs an ideally constant width) compared to the pre-optimised deposition behaviour.