Interpretable multi-objective machine learning with calibrated uncertainty for deployment-oriented prediction of defects and properties in polymer FFF

Fused filament fabrication (FFF) parts exhibit variability in defects (porosity, interlayer weakness, and warpage) and scatter in mechanical properties due to interacting process parameters and environment. Many predictive tools are black-box and lack calibrated uncertainty, limiting deployment. An interpretable, deployment-oriented framework is introduced that integrates physics-informed analytical synthesis, empirical anchoring, calibrated uncertainty quantification (UQ), and multi-objective optimization across four polymer archetypes: polylactic acid (PLA), acrylonitrile–butadiene–styrene (ABS), poly(ethylene terephthalate)-glycol (PETG), and thermoplastic polyurethane (TPU). A hybrid dataset combines synthetic samples (n = 400 per material) with literature-validated experimental records (n = 20 per material). Ensemble surrogates, gradient boosting, random forests, support-vector regression, and shallow neural networks were used to predict tensile strength, elastic modulus, and dimensional error; SHapley Additive exPlanations (SHAP) provide mechanism-consistent attributions; isotonic regression calibrates prediction intervals. Gradient boosting attains strong tabular performance (R² ≈ 0.98 for strength; R² ≈ 0.99 for dimensional error). Post-hoc calibration lifts empirical coverage from ∼ 60–70 % to ∼85–90 %, enabling accept/reprint gating. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) traces Pareto fronts, and a knee-point selector returns actionable trade-offs between strength, dimensional accuracy, and process time. Key contributions are: (i) a physics-informed synthesis regime with explicit failure-mode injection that broadens exposure to defect-prone states; (ii) calibrated UQ that converts ensemble variance into reliable error bounds; (iii) interpretable attributions revealing dominant, material-specific drivers (e.g., printer model, layer height, cooling); and (iv) knee solutions for each polymer. Although enclosed FFF platforms can be repeatable, residual uncertainty persists (lot variation, geometry-dependent thermal histories, cross-printer effects). The workflow is transparent and reproducible, and pre-registers criteria for future extension to direct energy deposition (DED) and related processes.