Zeng, Yu2026-04-302026-04-302026-04-302026-04-27https://hdl.handle.net/10012/23123High performance composite materials, particularly carbon fiber reinforced polymer (CFRP) composites based on unidirectional non-crimp fabric (UD-NCF) reinforcements, have gained widespread adoption in aerospace, automotive, and wind energy industries due to their exceptional specific stiffness, strength, and design flexibility. However, the mechanical performance of UD-NCF composites is strongly governed by complex manufacturing induced microstructural features and defects, including irregular fiber geometries, non-uniform fiber dispersion, in-plane fiber misalignment, and out-of-plane fiber crimping, which arise inherently from the fabric architecture and liquid resin infusion processes. These microstructural complexities introduce significant variability in effective mechanical properties, posing substantial challenges for accurate property prediction. The inherent complex multiscale structure of UD-NCF composites must be considered when predicting mechanical properties, where finite element (FE)-based approaches are preferred. However, the high computational cost of repeated high fidelity FE simulations across multiple length scales renders conventional multiscale approaches impractical for design optimization, sensitivity analysis, and uncertainty quantification. Addressing these limitations requires a computationally efficient, physically informed multiscale framework that explicitly incorporates realistic microstructural features while remaining tractable for engineering applications. The focus of this thesis was to develop a scalable multiscale computational framework integrating finite element (FE) homogenization and machine learning (ML) surrogate modelling to accurately predict the effective elastic properties and strain rate dependent inelastic behaviour of UD-NCF composites, while explicitly accounting for realistic microstructural features and manufacturing induced variability. The first objective was to develop a hierarchical dual scale FE homogenization framework explicitly incorporating key manufacturing induced microstructural features quantified through systematic microscopic characterization. Microscale representative volume elements (RVEs) captured fiber geometry, spatial dispersion, and resin distribution, whilst mesoscale models incorporated tow geometry, inter tow spacing variability, in-plane fiber misalignment, and out-of-plane fiber crimping. Microscale fiber distribution and diameter variability had negligible influence on mean elastic properties for sufficiently large RVEs, whilst RVE size significantly affected prediction variance. Microscale predictions showed strong agreement with Chamis' analytical formulations, with moderate discrepancies in shear properties at lower fiber volume fractions. At the mesoscale, in-plane misalignment exerted a pronounced influence on the longitudinal modulus and in-plane Poisson's ratio, whilst out-of-plane crimping had minimal effect at small angles. The framework predicted longitudinal and transverse elastic moduli within 1% of experimental measurements, with approximately 20% deviation in the in-plane shear modulus attributed to uncertainties in fiber shear properties and microstructural variability. The second objective was to develop an automated data generation pipeline and train ML surrogate models to establish quantitative elastic structure-property relationships for the UD-NCF composite material at both length scales. Neural network and XGBoost models trained on systematically generated FE datasets achieved mean R² values exceeding 0.97 at the microscale and 0.99 at the mesoscale, with mean absolute percentage errors consistently below 1.5%, reducing prediction times from hours to seconds. SHAP based interpretability analysis confirmed that the surrogate models captured physically meaningful structure property relationships consistent with established micromechanical theory. The third objective was to extend the multiscale framework to predict strain rate dependent inelastic behaviour of UD-NCF composite by integrating established physical constitutive formulations into the surrogate modelling workflow, addressing a key gap in data driven composite modelling where inelastic deformation behaviour has received comparatively limited attention. Neural network and XGBoost surrogate models were trained on FE RVE datasets incorporating systematic variations in microstructural parameters, including in-plane fiber misalignment, with the FE framework showing strong agreement with experimental inelastic and strain rate dependent responses, confirming its suitability as a data generation tool. An automated RVE generation and post-processing pipeline provided a scalable foundation for efficient dataset construction. Incorporating physical formulations, namely the Johnson Cook strain rate model, Hill's anisotropic yield criterion, and power law hardening, into the data engineering process substantially reduced training data requirements, constrained the solution space, and improved model generalizability, enabling accurate prediction of complex inelastic and strain rate dependent behaviour across a wide range of loading rates. Collectively, this work establishes a scalable, computationally efficient digital twin framework bridging microstructural variability and macroscopic composite performance. By explicitly accounting for manufacturing induced microstructural features and integrating high fidelity FE homogenization with data driven surrogate modelling, the developed methodology provides a practical tool for accelerating multiscale modelling, virtual prototyping, sensitivity analysis, and design optimization of UD-NCF composite systems under realistic manufacturing conditions, representing a significant step towards physics informed digital twins for next generation composite structural applications.enDoctoral Thesis