Wind Flow Prediction through Machine Learning
One of the most difficult to assess aspects of urban performance has traditionally been the wind flow around buildings, whether that is thought of as an indicator of pedestrian comfort and safety, natural ventilation or pollution concentration.
The difficulty of course stems from a combination of almost inifite number of “urban chunks” within a city and the computational and time constraints of a Computational Fluid Dynamics (CFD) simulation at the level of an urban neighbourhood. This has typically allowed the use of CFD simulations almost entirely as a validation step for the, already finalized, proposed design.
It is this that has made parametric optimization of wind flow around proposed developments practically infeasible, within the typical constraints of a real-life project timeline.
Machine Learning can help us change this. It can bring the world of CFD and wind flow assessment into the realm of iterative evaluation and optimization of design and equally important, can allow us to conduct even more complex studies of urban performance such as thermal comfort.
In this project we are using a Deep Learning model to predict CFD simulations in real time. A Generative Adversarial Neural Network has been trained with thousands of CFD simulations from the digital 3D model of the city of Vienna and is able to instantly predict any new design it is given.
The combined Deep Learning models for Solar Radiation and Wind Flow prediction are used in an interactive design framework that allows users to draw arbitrary geometries and get instant Solar Radiation and CFD results.