Neural networks and NARXs to replicate extrusion simulation in digital twins for fused filament fabrication

In this work we propose the use of nonlinear autoregressive models with exogenous variables (NARXs), powered by dynamic recurrent neural networks, to replicate the results of a previously developed, complex simulation of the extrusion process in fused filament fabrication. The NARXs predict extrusion rate of the extrudate and compression force acting on the filament with an average discrepancy of 0.12 % with respect to the original simulation, but at a fraction of the computational time (0.1 s of the NARX vs. 600 s needed by the original simulation to process the same one-minute time interval). In addition to illustrating how NARXs can be created to mimic an existing simulation, in this work we show how the NARXs can be physically connected to the sensors of a real FFF machine, thus creating an effective digital twin of the extrusion process, useful to support real-time decision making by an AI machine controller. Finally, we show how the implemented digital twin can be used for in-process monitoring, bringing as example the automated detection of an extrusion clogging event.