Poor indoor air quality contributes to poor respiratory health at an individual and public health level, with subsequent economic effects. These problems could be tackled by improving ventilation. While computational fluid dynamics simulations can model air flow in a room and identify ventilation issues, calculations are computationally expensive, and changing parameters can yield little to no improvement or even worsening of ventilation. A required iterative process is less than ideal for static environments and too expensive for widespread application. To approach this problem, the aim of this project was to generate sufficiently rapid simulation predictions that could incorporate changes inside the rooms environment, an approach which requires a focus on velocity patterns. Since spee...[ Read more ]

Poor indoor air quality contributes to poor respiratory health at an individual and public health level, with subsequent economic effects. These problems could be tackled by improving ventilation. While computational fluid dynamics simulations can model air flow in a room and identify ventilation issues, calculations are computationally expensive, and changing parameters can yield little to no improvement or even worsening of ventilation. A required iterative process is less than ideal for static environments and too expensive for widespread application. To approach this problem, the aim of this project was to generate sufficiently rapid simulation predictions that could incorporate changes inside the rooms environment, an approach which requires a focus on velocity patterns. Since speeding up conventional methods is found to be inadequate, artificial intelligence was utilized as an enabling technology. A two dimensional parameterized interface was chosen to speed up dataset generation and simplify data processing. Out-of-sample fluid flows could be predicted with an average coefficient of multiple correlation of 0.5. The out-of-sample mean average percentage error strongly indicated a lack of features for machine learning, with the flow components in the Y direction being just below 5%, while that of the X components were approximately 57%. Overall, this approach emphasized improvements in calculation time over accuracy. The average calculation time reduced from approximately 40 hours for conventional calculations, to around 4 seconds using a trained network. Future investigations will optimize the methodology to improve both accuracy and computation time further.