AI-based prediction and numerical analysis of mechanical and microstructural parameters of carbon fiber reinforced 3D printed composites

This study introduces a hybrid CNN-DNN-ensemble FEA framework for predicting mechanical and microstructural properties of 3D-printed carbon fiber-reinforced thermoplastic (CFRTP) composites via synthetic data generation. Representative volume elements (RVEs) with randomized fiber positions/orientations (volume fraction: 10–30%) and porosity (<5%) were created using DCGANs trained on CT datasets. Synthetic RVEs were meshed and subjected to uniaxial tensile loading in ABAQUS/COMSOL, yielding mean tensile strength of 582.4 MPa and elastic modulus of 48.2 GPa. Microstructural features (porosity, fiber orientation tensors, voids) were extracted via CNN-based segmentation and used in multitask DNNs enhanced with CNNs, Random Forest, and XGBoost; physics-informed regularization enforced constitutive laws. Multitask DNN and ensemble predictions achieved R² > 0.989, RMSE < 6.27 MPa, and errors < 2.1% against FEA for mechanical properties; microstructural errors were < 3.2%. SHAP analysis identified fiber volume fraction (importance: 0.314) as the key driver of tensile strength. Parametric studies revealed optimal fiber volume fraction (~60%) and alignment (<15°), with multitask DNN–FEA discrepancies < 2.4%. ASTM D3039 validation showed deviations < 1.1% (strength) and < 0.7% (modulus). This framework addresses data scarcity in additive manufacturing, enabling rapid, physics-constrained multiscale analysis for aerospace and automotive design optimization.