Predicting the mechanical behavior in FDM printing of biopolymers through boosting artificial neural networks

Additive Manufacturing (AM), and in particular Fused Deposition Modeling (FDM), offers unprecedented flexibility in fabricating complex geometries. However, the mechanical behavior of FDM-produced specimens is intricately dependent on a range of printing parameters and testing conditions. In this study, a novel machine learning framework designed to map these process parameters to the resultant mechanical properties of specimens fabricated using polylactic acid (PLA), both in its neat form and when reinforced with natural fibers (flax and wood flour) is presented. The proposed approach leverages Artificial Neural Networks (ANNs), primarily utilizing a Gaussian/Gumbel-Gaussian Multi-Layer Perceptron (MLP) architecture enhanced by an encoder module derived from a Variational Autoencoder (VAE) trained on the input dataset to effectively compress and represent high-dimensional process data. To further bolster predictive performance, a boosting neural network scheme based on this architecture and a kernel based truncated Taylor expansion are integrated. Experimental validation reveals that the presented model achieves high accuracy, as demonstrated by a low Continuous Ranked Probability Score Normalized (CRPSN) of 0.072 and an improvement with respect to a mean baseline (Skill) of + 67 %.