I really just wanted to share here some of my current work that has been focusing lately on a connection between my domain expertise, environmental performance-based design and optimization, and ML/DL methods and models. This work has been trying to assess the potential of various Deep Learning models, mainly generative models, in streamlining urban environental performance studies and urban design itself.
I know that @jeremy and @rachel have been making a point to promote applications of DL that have the potential to change the world for the better. This is the main reason why I thought that such a post would not feel out of place in this forum (I do hope I’m right). I am convinced that under heavy urbanization, and the pressure of Climate Change, this use case has immense potential in changing how we think, assess, design, and inhabit cities and I would love it if more people get involved with it. I am hoping that this post might spur some interest.
As I said the practical side of this has been focused on generative models. I have been able to train a few different architectures, going through most of the state-of-the-art VAE and GAN models, in an effort to predict Urban Environmental Performance (UEP). After a few tests I have somehow settled on image-to-image translation architectures such as [bi]cycleGAN and MUNIT, which seem to do a great job in producing good accuracy results. These models are trained on both synthetic datasets of urban configurations that I procedurally and parametrically produce as well as close-to-real-life data (I’m currently using Microsoft’s US buildings dataset: https://github.com/Microsoft/USBuildingFootprints).
The input data are grayscale heightmaps of urban areas, which allows me to ‘communicate to the DL model’ the 3D massing dimension of the input even when I am using top-down, 2d images. The outputs are the same inputs, but now colored by environmental performance. So far, I have been using solar radiation studies as a proxy, mostly because it is a very quick simulation even at urban scales, which cuts down the time needed to generate the datasets.
The models, so far, seem promising and appear to have a good ‘understanding’ of the correlation between building shapes, heights, and environmental performance. I have tried to test their capacity to generalize by feeding them either parts of the dataset that they haven’t seen before or even unrelated urban configurations. Results are satisfactory, even at these early stages, although I might have a bit of excitement bias.
I’m currently working on how to better serve these models in order to develop a design tool that will allow near real-time performance assessment. Something like this, especially when the focus turns to important studies that typically require heavy computation and time (a thermal comfort study for example requires 1 week at this scale), and don’t really scale well with size and number of alternatives, has the potential to completely transform design processes.
There are a lot of things yet to test and try out, e.g. input multiple slices at different heights or multi-domain translation of different performance maps, and so on. There are also, I’m sure, a lot of ideas that I haven’t come up with yet or I never will. Part of why I’m posting this is the chance that some of you might contribute in this.
I am currently preparing a blog that will go through the whole process and hopefully lower the barrier for people without the specific domain expertise to get involved. Unfortunately, it is not ready yet. The best I can share right now is a series of LinkedIn posts that I have written. They should give a general idea of the problem setting and preliminary results:
If you find this interesting, don’t hesitate to reach out! I’m more than willing to help you get involved and working on this!