Architected materials offer unprecedented control over macroscopic mechanical and thermal behavior through tailored internal geometry. Despite major advances in additive manufacturing [1], deriving generalizable and interpretable structure–property relationships remains challenging due to the high dimensionality of geometric representations [2] and the prevalence of architecture-specific, image-based learning approaches. This work introduces a geometry-driven machine learning framework for predicting effective mechanical and thermal properties of architected materials directly from unit-cell design descriptors. Rather than relying on topology-dependent encodings, the approach employs a compact set of physically interpretable geometric features extracted from CAD-level representations, capturing volumetric content, internal surface characteristics, inertial resistance, and spatial heterogeneity. A comprehensive dataset is generated via voxel-based finite element homogenization, spanning elastic stiffness, shear response, yield initiation, and thermal conductivity across a broad range of relative densities and geometric configurations. Ensemble learning models are trained on these descriptors and achieve high predictive accuracy for all target properties. Model interpretability is addressed through SHapley-based explainability analysis, enabling systematic ranking of feature importance and identification of dominant geometric mechanisms governing material response. The results reveal hierarchies of geometric influence, with the volumetric fraction to control global scaling. Remarkably, feature reductions techniques reveal that the mechanical performance of well-known sheet-based metamaterials can be predicted by a limited subset of shape descriptors. By linking CAD-level geometry, finite element homogenization, and interpretable machine learning, the proposed framework provides a scalable and data-efficient pathway for performance prediction and rational design of architected materials, with direct applicability across diverse cellular and lattice-based systems.