Operator Learning for Indoor Air Dynamics: Modeling and Applications

Yuexin Bian, Yuanyuan Shi

Keywords: Data-driven modeling; Indoor air dynamics; Operator learning; Computational Fluid Dynamics; HVAC control;

Abstract: Indoor air quality plays a critical role in ensuring occupant health, comfort, and energy-efficient building operation. Accurate prediction of indoor airflow and precise boundary-level HVAC control can significantly enhance building performance. While Computational Fluid Dynamics (CFD) offers high-fidelity modeling, its computational cost makes it impractical for real-time applications. To address this, we propose an ensemble neural operator transformer (ENOT) that predicts the spatiotemporal evolution of indoor CO$_2$ levels, achieving a 250,000× speed-up over traditional CFD simulations. Our contributions include a high-fidelity CFD-based dataset, a simulation pipeline for realistic indoor air modeling, and an ensemble neural operator learning framework for accurate, real-time inference. We further outline future directions in data-driven model-based HVAC control, bridging the gap between high-fidelity simulation and intelligent building management. Our code and data are publicly available at https://huggingface.co/datasets/alwaysbyx/Bear-CFD-dataset.

Submission Number: 3