Machine learning-based prediction and analysis of dimensional accuracy in fused deposition modelling of Nylon-Aramid fibre composite parts

Fused Deposition Modelling (FDM) of fibre‑reinforced composites offers enhanced mechanical properties but introduces thermal and rheological interactions that compromise dimensional accuracy. Nylon–Aramid short‑fibre composites are promising, yet their dimensional behavior under varying process parameters remains uncharacterized. Most machine learning (ML) studies on FDM accuracy rely on in‑sample metrics or k‑fold cross‑validation, leading to overoptimistic generalization estimates. A Taguchi L 27 array varied five FDM parameters (layer thickness 0.20–0.40 mm, print temperature 280–300 °C, raster angle 0–90°, infill pattern, infill density 70–90%) for a Nylon‑Aramid (98/2 wt%) composite. Dimensional deviations (length, width, thickness, diameter) were measured. Five ML models, Gaussian Process Regression (GPR), Support Vector Regression (SVR), Random Forest (RF), XGBoost, and Artificial Neural Network (ANN), were trained and evaluated under leave‑one‑out cross‑validation (LOO‑CV). XGBoost achieved the best generalization (mean RMSE = 0.02658, mean MAPE = 37.65%). Feature importance showed print temperature as dominant for length deviation, while infill density drove width, thickness, and diameter deviations, departing from Taguchi rankings. Parametric effect curves revealed a non‑monotonic temperature dependence (optimal near 285–292 °C) for length deviation. The optimal combined setting (0.20 mm, 290 °C, 45°, triangular infill, 70% density) was Experiment 4, which had the lowest combined dimensional accuracy index. This first ML‑based dimensional accuracy study for Nylon‑Aramid FDM composites shows that XGBoost with LOO‑CV provides robust generalization on small DOE datasets. The identified optima offer actionable guidance for process engineers.