Agent-based pedestrian flow simulations are often involved in planning large events and optimizing the operation of critical infrastructures. Although they show great potential to improve reactive measures in crisis situations via what-if analysis, their computational complexity limits their applicability in real-time data driven scenarios. To solve this issue Nellinger et al. have introduced a hybrid navigation model. The hybrid pedestrian navigation model combines global route planning with local, agent-based motion provide a strong foundation for scalable digital twin simulations of large infrastructures. Building up on this model we propose the use of geometry splitting to further improve computational efficiency in large-scale environments. Therefore, the simulation domain is partitioned along natural architectural boundaries such as rooms, corridors, and enclosing walls or even whole buildings, dependent of the use case. Each subdomain contains its local geometry, while global navigation remains coordinated through a shared graph representation. Agent-movements are calculated within the subdomains allowing the parallelization of the computational expensive parts of the simulation. The singular instances communicate only in case of agents leaving one subdomain entering the other. Using representative infrastructure scenarios with varying geometric complexity and pedestrian densities, we analyze the impact of room-based domain splitting on computation time and memory usage and demonstrate that geometry splitting, combined with hybrid navigation, is an effective approach for scalable pedestrian flow simulations and enhance their applicability in crisis scenarios.