AI-driven process optimization uses particle swarm optimization (PSO), reinforcement learning, and CNN-enabled calibration to reduce cost, time, and scrap rates in FFF/FDM printing. Key achievements include:

• PSO Experimenter — reduced build cost from $24.62 to $3.22 (87% reduction) and parameter tuning time by 97%
• Autonomous calibration — CNN defect detection with simulated annealing achieved 0.047mm average deviation
• Reinforcement learning agents — iteratively learn optimal nozzle temperatures, raster angles, and speeds for complex geometries

Machine learning models rank sustainable polymers and composites by balancing mechanical performance with environmental impact. Methods include:

• XGBoost regression — lowest RMSE among six regressors for tensile strength prediction
• ANN-GA hybrids — identified optimal parameters (0.28mm layer thickness, 34% infill, 222°C nozzle)
• Pearson heatmaps — reveal how shell thickness, nozzle temperature, and layer height drive tensile strength

Supervised learning and hybrid neuro-fuzzy systems forecast mechanical properties from print parameters, enabling digital twins for real-time process control:

• ANN + fuzzy logic — predicted tensile strength with 2.21–3.29% error; infill density identified as dominant variable
• Feed-forward perceptrons — achieved R² ≈ 0.994 across tensile strength, surface roughness, build time, and material consumption
• Digital twin estimators — MLP and gradient boosting provide live inputs for cyber-physical sustainability monitoring

AI-driven generative design and topology optimization reduce design iteration time by 30–50%, part weight by 10–50%, and manufacturing costs by 6–20%. Approaches include:

• VAEs and GANs — generative models for engineering design considering performance and aesthetics
• CNN-based topology optimization — Pyramid U-Net architecture with multigrid techniques achieves 99% accuracy
• ANFIS for TPMS structures — predicts energy absorption and peak load for lattice structures
• Custom splint optimization — 149% improved heat dissipation, 42.6% weight reduction via FDM

Computer vision and deep learning enable real-time anomaly detection, achieving over 80% defect detection rates:

• SSD + VGG16 — trained on 2,500 frames; precision 0.44, recall 0.69 at IoU 0.4; deployed on Raspberry Pi
• XceptionTime — 95%+ accuracy classifying environmental sensor data (temperature, humidity, pressure)
• Laser point-cloud + CNN — 90% anomaly classification with PID control maintaining 0.1% layer height deviation
• Digital twins — lightweight CNN for FDM fault classification with F1-score of 0.998

ML enables energy monitoring, life cycle assessment (LCA), and waste reduction for sustainable manufacturing:

• LSTM energy classification — 98.2% accuracy classifying print/standby/preheat stages across PLA, ABS, PETG
• ANN-optimized exoskeleton — 30.9× waste reduction, enabling one free print per 6.67 prints via recycling
• Live LCA — cyber-physical systems with Taguchi L-9 arrays optimize environmental impact in real-time