Multiphysics cardiac simulations coupling fluid-structure interaction and electrophysiology (FSEI) [1] have demonstrated the capability to reproduce complex cardiac dynamics. However, most existing stud- ies are limited to single, often patient-specific, anatomies, whereas population-level variability must be incorporated to enable robust numerical modeling, uncertainty quantification, and in-silico trials. This requires the availability of virtual cohorts of anatomically consistent cardiac geometries. Traditional seg- mentation approaches, however, proved incapable of building such robust, accurate and rich cohorts due to several acquisition and operator-dependent limitations [2]. Even though Deep learning-based methods have partially overcome these difficulties, this is still an open problem. As a consequence, the geome- tries resulting from these large-scale segmentation databases often exhibit mesh defects and topological inconsistencies, such as holes and non-manifold regions, which are not admissible for multiphysics sim- ulations, especially with hemodynamics. In this work, we propose a semi-automatic pipeline to generate simulation-ready cardiac meshes directly from CT scans. The method integrates machine learning-based segmentation, principal component analysis [3], and in-house geometric processing libraries to ensure watertight surfaces and mesh quality compatible with multiphysics solvers. The pipeline was applied to a dataset of 60 cardiac CT scans, enabling the construction of a statistical shape model to quantify anatomi- cal variability. This model is further exploited for data augmentation and the generation of virtual cardiac cohorts. The generated geometries have been successfully employed in multiphysics cardiac simulations starting from healthy subjects, highlighting the potential of the proposed framework for population-based cardiac modeling and large-scale in-silico studies.