OmniFlow: A Framework for Generalizable Surrogates for Real-Time Airflow Simulation and Control in Unseen Indoor Environments

High-fidelity CFD simulations are essential for predicting airborne contaminant transport, yet their computational cost prohibits use in real-time building control. Data-driven surrogates offer the necessary speed but have historically been limited by an inability to generalize to unseen geometries, a critical barrier to widespread adoption. This paper introduces OmniFlow, a complete a framework that addresses this challenge. We present a methodology where a surrogate is trained exclusively on a diverse, procedurally generated dataset of indoor environments. The training is guided by a hybrid loss function that incorporates control-aware advection and gradient-matching physics terms, enabling the model to learn the underlying principles of fluid dynamics. This approach allows our surrogate to perform robust zero-shot inference on complex, real-world scanned geometries by accelerating the iterative pressure solve within a larger physics simulation. An ablation study confirms our hybrid loss is the critical component for achieving this physical fidelity. We demonstrate the surrogate's capability by integrating it into a gradient-based Model Predictive Control (MPC) framework to dynamically optimize HVAC airflow and direction following a cough event. In realistic, scanned environments, our MPC significantly improves contaminant clearance over strong baselines while offering substantial energy savings. This work provides a complete, validated framework for developing and deploying the next generation of intelligent, health-aware control systems.