Comparison of the performances of Statistical and Artificial Neural Network models in the prediction of geometry and density of PLA/wood biocomposite cubes manufactured by FDM

The present study investigates the impact of scanning speed, printing temperature, and layer height on the density, dimensions, errors of parallelism, and surface finish of cubical specimens made of PLA/wood biocomposite and manufactured by Fused Deposition Modelling (FDM). The study examined 64 specimens, each produced with a unique set of process parameters. The Response Surface Methodology (RSM) was employed to evaluate the effects of process parameters on the examined properties of the manufactured cubes. RSM analysis revealed the statistical significance of direct proportion between the layer height, printing temperature, and x-and y-dimensions of the manufactured specimens (with P-values of 0, 0, 0.002, and 0, respectively). Also, the scanning speed and error of parallelism in z-oriented faces were statistically correlated (with a P-value of 0.035). For layer height and cube density, an indirect proportion was observed (with a P-value of 0). Compared to the regression model, ANN exhibited better performance at process parameters effect evaluation. The worse performance of regression models can be attributed to their limited capacity to represent non-linear relationships, while ANN models can capture the complex non-linear nature of the process, leading to better performances (R2 close to 100%). An evaluation of the defects in the specimens was carried out using the go/no-go diagram.