| dc.description.abstract | Short-fiber-reinforced thermoplastics (SFTs) manufactured via additive manufacturing–compression molding (AM-CM) provide a unique combination of highly aligned fibers (>80%) and low residual porosity (<2%). These characteristics make them promising candidates for fatigue-critical structural applications in aerospace, automotive, and wind energy systems. Despite growing industrial adoption, rapid methodologies for fatigue characterization and computational frameworks that link process-induced microstructures to cycle-dependent performance remain limited. This dissertation addresses these challenges by developing an experimentally validated, multiscale framework that integrates passive infrared thermography (IRT), micro-computed tomography (μ-CT), finite-element (FE) homogenization, and machine learning surrogates for carbon fiber/acrylonitrile butadiene styrene (CF-ABS) SFTs processed via AM-CM.
First, on the experimental side, a multi-indicator IRT cyclic step loading protocol is developed and validated as an accelerated alternative to conventional stress-life (S-N) testing. By monitoring four thermal indicators, the proposed protocol predicts the high-cycle fatigue strength (HCFS) within ±4% of conventional S-N results using only three specimens in five hours, reducing testing time by approximately 96X. Systematic parametric investigations further establish that HCFS increases by 17.3% as loading frequency rises from 5 to 20 Hz due to suppression of viscoelastic creep rather than thermal effects, and decreases by 19.2% as the stress ratio increases from R = 0.10 to R = 0.75. Fatigue fracture entropy (FFE) is further demonstrated to remain invariant across the tested frequency range, providing a thermodynamic basis for frequency-corrected fatigue limit estimation.
Second, porosity is investigated as a fatigue-sensitive damage state variable through interrupted fatigue experiments combined with stage-wise μ-CT imaging. Specimens spanning 1.41%-2.86% initial void content exhibit a 24.4% reduction in HCFS. Above the fatigue limit, void fraction increases by 232%-370% due to localized pore growth near fiber ends and matrix-rich regions. Stage-resolved FE homogenization based on experimentally derived microstructural distributions predicts a 2%-4% relative stiffness degradation, consistent with experimental measurements.
Third, to bridge microstructure and mechanical response, μ-CT fiber arrangements are encoded as topology-aware weighted interaction graphs and processed by a hybrid graph neural network alongside a long short-term memory (GNN–LSTM) surrogate trained on nonlinear finite element (FE) homogenization data. The surrogate accurately predicts nonlinear stress-strain trajectories (R2 = 0.993) with a mean error below 4% across several microstructures while assembling coupon-scale responses in just 60 seconds, representing a 100X computational speedup over direct FE simulations. This enables rapid evaluation of microstructure-dependent constitutive behavior that would otherwise require computationally intensive nonlinear FE analyses.
Finally, the framework is extended to macroscale fatigue life prediction via a coupled thermomechanical transient implementation of matrix constitutivity that incorporates viscoelasticity and temperature-dependent stiffness degradation within a continuum damage mechanics framework. A cycle-jump acceleration algorithm reduces computational cost by 100X while maintaining less than 2% error in damage evolution.
Together, these contributions establish an integrated pathway from accelerated experimental characterization to an accelerated surrogate-based multiscale fatigue life prediction, thereby enabling more efficient qualification, design, and deployment of additively manufactured thermoplastic composites for fatigue-critical structural applications. | en_US |
