Module 8 Questions:
Part 1:
The main problem that the speakers brought up was that building energy simulation is too slow and specialized to be useful during early conceptual design, when design changes are cheapest and most impactful. They bring up using a machine learning surrogate model to predict Energy Use Intensity, trained on synthetic data they generated using Dynamo, Generative Design in Revit, and Autodesk Insight.
What surprised me is how little usable training data exists for buildings despite how many plans and models are out there. The problem they had for this is that real building data is too inconsistent in format and quality to reliably pull data from. As documentation becomes more standardized I think this gap will close and surrogate models could eventually be trained on real buildings instead.
I'm not sure how confident I am in the surrogate model currently, especially for a more specialized or complex building. This method is good for early preliminary understanding, but I would probably feel most confident about using this method for smaller buildings in a non-seismic zone that could be more characteristic of a controlled environment since the synthetic data is more likely to be representative.
Part 2:
One place I experienced some friction was during the image pixel geometry assignment where we had to figure out how many pixels were necessary to capture the image properly without over-defining it or underdefining it. Too many pixels were unnecesary and demanding, too few wouldn’t capture the image. There is likely a better defined point where the image is understandable and not overdesigned. This is something that AI-aiding would be helpful for. An AI model could learn the relationship between pixel size and visual perception and find a point that works best for each different image, rather than using guess and check.
Part 3:
Swarm (Thornton Tomasetti): ML tool that optimizes structural configurations and sizing in the early design phase. It helps with optimizing and introducing creative solutions earlier than traditional workflows. This is very helpful for structural engineers when brainstorming potential layouts, especially under architect wants. This doesn’t directly help with what we did this quarter, but potentially in the workforce it would be intereseting to look into.
Autodesk Forma: Given the parameters, it allows the user to conduct environmental analysis of a building with outputs to understand daylight, wind, noise, and embodied carbon using AI. It can also generate and evaluate many layouts based on the defined parameters. This would be very useful for owners who want to meet certain LEED or WELL certification parameters because it would directly help them understand how different building designs will change their building rating. This coul be helpful for us when defining the sunlight parameters.
Hypar: The user can define the parameters of a building as in depth or abstract as they want and it will generate different potential geometries and BIM outputs. It is used alongside Dynamo or Grasshopper and is not meant to be an alternative. This would be very helpful as a tool to help understand why my software isnt working and would have been very useful in the building parameterization we did to help run different geometries.