Understanding and predicting airflows in enclosed environments not only lowers physical dangers in extreme situations but also has a wider and long-term impact on safety, thermal comfort and energy efficiency. Current research uses Computational Fluid Dynamics (CFD) methods for airflow simulations. It has been used for optimising building designing or improving existing ones. However, few people model airflows with crowds together. Coupling two complex systems with different dynamics imposes a great challenge.
The project aims to combine two systems together into a single framework. Since both are computationally expensive, data-driven methods will be leveraged, especially cutting-edge machine/deep learning methods. Deep learning (DL) has been successful in many problems. Recently, there is a surge in using supervised DL for accelerating CFD in computer graphics and crowd prediction. However, the effort has only been made in two separate fields.
The project will first extend a DL-based fluid simulator for airflow simulation from fixed boundary conditions (as it is in the literature) to dynamic boundary conditions to accommodate crowds. Next, a data-driven crowd predictor will be incorporated. Finally, a single framework that can predict crowd movements and the corresponding airflows will be proposed.
The major outcome of the project is a framework that can interactively simulate airflows with crowds. To this end, it will identify the performance bottlenecks in both CFD and crowd simulation in relevant scenarios. It will also propose new methods that can leverage big data and accelerate simulation and prediction.
The specific form of outcomes includes publications of top venues, an open-source project including data and code, and potentially plug-ins for existing pipeline and software platforms.
The project is currently accepting PhD applications every year.