Hybrid Machine Learning Framework for Predicting the Biomechanical Performance of FDM-Manufactured PEEK Bone Plates

Abstract

This study investigates the optimization and prediction of the mechanical properties of angular stable bone plates fabricated by fused deposition modeling (FDM) using polyether ether ketone (PEEK). A full-factorial experimental design comprising 81 parameter combinations was used to evaluate the effects of layer thickness (100, 200, and 300 μm), infill pattern (gyroid, cubic, and triangles), printing speed (20, 40, and 60 mm/s), and infill density (30%, 60%, and 90%) on tensile strength, flexural strength, and Shore D hardness. A sparse autoencoder (SAE)-based feature representation was combined with support vector regression (SVR), random forest (RF), and artificial neural network models (ANNs), whose hyperparameters were tuned using grid search optimization (GSO). Analysis of variance (ANOVA) identified infill density as the most influential processing parameter for all three mechanical responses. Among the evaluated prediction models, the optimized SVR achieved the best overall performance, with R2 values of 0.9653, 0.9851, and 0.9987 for tensile strength, flexural strength, and Shore D hardness, respectively. An additional ablation analysis showed that the contribution of the SAE transformation was model- and response-dependent, with the clearest improvements observed for SVR in tensile- and flexural-strength prediction. These results affirm the efficacy of machine learning models─particularly SVR with SAE-based nonlinear feature transformation and hyperparameter optimization─in accurately predicting and optimizing the mechanical behavior of components produced by additive manufacturing (AM), which is directly relevant to biomedical device fabrication.