Generative Machine Learning Models for Airflow Prediction of Architectural Spaces

Buildings have been identified as one of the biggest contributors of negative environmental impacts worldwide, more specifically energy usage due to the use of air conditioning and mechanical ventilation. Strategies such as cross-ventilation have become a reliable alternative to diminish some of these effects. However, designing for cross ventilation is no easy feat, as it requires architects and designers to study in detail the building context, overall massing design and building enclosure to maximize airflow potential. On the other hand, Computer Fluid Dynamics (CFD) airflow simulations are not used as often in architectural settings primarily due to time constraints, lack of performance metrics and quality assurance. The proper use of CFD airflow simulations involves a complex setup and run-time process, due to the large mathematical calculations involved.This study aims to apply existing generative machine learning algorithms to compute CFD wind velocity simulations to significantly shorter run times while maintaining a relatively high accuracy level, during the initial design stages. To test the proposed hypothesis, multiple machine learning models were created, trained, and tested. The evaluation metrics for these models consisted of using different image similarity methods to compare the images produced by the machine learning model to their CFD engine counterparts. The results obtained indicate that GAN application for CFD airflow predictions can produce acceptable results showing a significant run time difference of over a minute between the CFD simulation and the machine learning model. Having evaluated and proven this study as a proof of concept, this can set the precedents for further research on the use of CFD airflow simulations and machine learning within architectural practice. Allowing architects and designers to incorporate the use of CFD airflow simulations within their workflows.