This project proposed and evaluated
a conditional generative
adversarial network surrogate
model that accounts for the contribution
of urban morphologies to
pedestrian wind flow conditions.
The validity of the approach was
demonstrated at a fraction of
the time that would be required
to perform the equivalent conventional
simulation. Variations
in dataset encoding techniques,
image resolutions and geometric
diversity of the training set
were explored to identify the key
parameters affecting model’s
accuracy and suitability. The model was
deployed within a CAD environment for
interactive near real-time performance
feedback during iterative design.