Cook, Brent; Gainza, Adryel; Portocarrero, Bryan; Mujawar, Mubarak; Kamat, Vivek; Ewing, Dan; Upadhyay, Himanshu; Bhansali, Shekhar. (2026).听.听Scientific Reports, 16(1), 29160.听
This study investigates physics-informed neural networks (PINNs) for reconstructing the position of a magnetic-rail nScrypt printer extruder from accelerometer measurements. The researchers first developed a two-body physics model of the extruder and then incorporated its equations of motion into neural networks in two ways: directly within the network鈥檚 activation layers or as constraints during model training. The activation-based approach produced a more detailed fit to the position signal, while the constraint-based approach better captured its average trend and variability. The two approaches were also combined and evaluated using root-mean-square error and standard deviation. Although the combined model showed low variability during training, errors varied substantially on test samples. Overall, the study explores how incorporating physical equations into neural networks can support sensor-based signal reconstruction and digital twinning of a 3D printer extruder.
Fig 1
The printer head of mass聽M聽and nozzle mass聽m聽diagram with position and deflection coordinates.