Predicting Linear Dimensional Accuracy of Material Extrusion Parts in Dependence of Process Parameters Using Neural Networks Optimized by an Evolutionary Algorithm

Abstract

Material extrusion (MEX) enables the economical manufacturing of complex parts and small lot sizes. The quality of the additive manufactured parts is significantly influenced by process parameters, which are defined beforehand. However, the resulting part quality is often unknown, leading to a reduced applicability of the process and relatively high safety factors for the process parameters to ensure a certain quality. This results in long print times and high material consumption. This article aimed for an accurate prediction of the linear dimensional accuracy in X, Y, and Z direction of parts manufactured with MEX. A neural network (NN) was used with a hyperparameter tuning based on an evolutionary algorithm. The developed NN achieved a mean absolute percentage error (MAPE) of 1.3% or lower for X and Y direction and 2.3% for Z direction by testing random parameter combinations (parameter sets). Moreover, when dealing with random and untrained interval lengths, the NN achieved a MAPE of 0.6% for X and Y direction and 3.3% for Z direction. The results show that the NN model achieved a more accurate and robust prediction compared with a multiple linear regression. The performed research fills an existing gap by developing a powerful NN model that enables the accurate prediction of linear dimensional accuracy based on used process parameters. The implications for practice are significant, as these prediction models can be readily used to improve the parameter settings for MEX and ensure that the desired accuracy levels are met. With further exploration of additional dimensional features and advances in data sharing techniques, the findings pave the way for future research to push the boundaries of accurate dimensional prediction in the field.