Taguchi-based Multi-parameter Optimization and Machine Learning Prediction of Mechanical Properties in FDM-Fabricated Polymer Composites

Abstract

The mechanical performance of parts produced via fused deposition modeling (FDM) is highly sensitive to various process parameters and material configurations, yet there remains a lack of comprehensive, predictive frameworks that link these settings to resulting mechanical properties, especially in multi-material systems. Addressing this gap is critical to advancing design and manufacturing practices in additive manufacturing. This study presents an experimental and data-driven investigation into the mechanical and morphological performance of polymer-based composites fabricated via FDM. Three different material configurations-pure PLA, carbon fiber-reinforced PLA (Cf/PLA), and sandwich composites-were evaluated under 27 distinct parameter combinations defined by a Taguchi L27 orthogonal array. Key process parameters included material type (MT), infill density (ID), printing speed (PS), infill pattern (IP), and wall thickness (WT). Tensile and compressive strength tests revealed that Cf/PLA exhibited the highest mechanical performance, reaching a maximum tensile strength (TS) of 48.89 MPa and compressive strength (CS) of 85.13 MPa. Sandwich specimens demonstrated enhanced compressive capabilities (up to 63.28 MPa) compared to pure PLA (max CS: 46.72 MPa), while maintaining moderate tensile strength (max TS: 42.34 MPa). Analysis of variance (ANOVA) indicated that MT (p = 0.004) and WT (p = 0.007) were the most statistically significant factors affecting tensile strength, while MT (p < 0.001) and ID (p < 0.001) dominated compressive strength outcomes. Scanning electron microscopy (SEM) of fractured surfaces revealed notable differences in interlayer bonding, fiber pull-out, and delamination among materials. Machine learning algorithms, including artificial neural network (ANN) and support vector regression (SVR), were implemented to improve the predictive accuracy of tensile and compressive strength estimations using the selected process parameters. The ANN outperformed other models, achieving R-2 = 0.9961 for tensile and R-2 = 0.9954 for compressive strength predictions, with MAPE values below 12%. Validation tests on unseen data confirmed the model's robustness, with ANN yielding the lowest average error across all cases. This study demonstrates that combining experimental design, SEM characterization, and AI-based modeling provides a powerful framework for optimizing FDM parameters. The findings offer valuable insights for developing structurally efficient and mechanically enhanced polymer composites in additive manufacturing.