Neural network-based modeling of FFF process for PET-G: Evaluating MLPNN and RBFNN performance in mechanical property prediction

Abstract

PET-G is a versatile thermoplastic resistant to impact loading, heat, and reactivity with solvents, and witnesses wide use in the Display and Signage, Packaging, Automotive, Electronics, and Medical Industries. Customized production using conventional techniques is expensive due to the high initial setup costs associated, in such cases 3D printing may be a better choice, but predicting its behavior during 3D printing remains challenging. Therefore, the current study aims to find effective process modeling techniques to achieve desired responses. The comparison of modeling techniques like MLPNN and RBFNN has been carried out to predict the mechanical properties at selected process parameters of FFF to significantly improve the mechanical property predictions. The MLP, which had 5 input dimensions and 20 hidden units, showed a CF of 0.99918 and specificity 1. Its MSE was 2.0751. On the other hand, the RBF network, with 5 input dimensions and 20 centers had a considerably lower MSE of 0.004589 but slightly lower CF values (0.92188 and 0.93393). These findings highlight that MLP excels in precision while RBF demonstrates the accuracy of the models for predicting mechanical properties.