Introduction: Fixed-form hyperelastic constitutive models struggle to represent the heterogeneous nature of biological tissues. To address this limitation, we propose a microstructure-driven, machine learning (ML)-based framework for constitutive modeling that directly learns macroscopic behavior from 3D finite element Representative Volume Elements (RVEs) [1]. The approach combines microfocus Computed Tomography (μCT) with Constitutive Artificial Neural Networks (CANNs) [2], enabling data-driven homogenization of arterial microstructures. Methods: Microstructural features of the medial layer of porcine descending thoracic aorta were quantified via μCT and used to generate RVEs incorporating elastic lamellae, elastin fibers, collagen fibers, and smooth muscle cells. Virtual biaxial tensile tests produced stress–stretch datasets reflecting microstructural variability. These datasets were used to train CANNs that embed hyperelastic structure, ensuring frame indifference and thermodynamic consistency. To account for microstructural variability, learned constitutive responses were clustered, enabling uncertainty-aware characterization. The framework was further applied to infrarenal arteries to assess adaptability across anatomical regions. Results: The trained CANNs accurately reproduced RVE-derived stress–stretch responses across loading modes. Clustering revealed distinct mechanical phenotypes associated with underlying microstructural configurations, providing a probabilistic characterization of tissue behavior. The learned strain-energy functions identified dominant anisotropic quadratic fifth-invariant contributions, recovering functional forms consistent with prior data-driven studies in, e.g., ovine pulmonary artery [3]. Application to infrarenal RVEs demonstrated region-specific constitutive models distinct from thoracic descending aorta behavior. Conclusions: This study presents a ML-driven homogenization framework that links arterial microstructure to macroscopic constitutive behavior. By integrating microstructure-informed RVEs with CANN modeling, the approach enables interpretable, variability-aware material representations. The framework is generalizable across tissues and species and provides a uncertainty-aware, ML-based constitutive modeling pathway for large-scale biomechanical simulations.