Module 8 Questions:
Option A - The technical pipeline: Using Generative Design and Machine Learning for Faster Analysis Feedback (AU 2020)
This talk explores a workflow to achieve early-stage analysis feedback for design iteration. They focus on early stage design because that’s when teams have a highest potential to make climate-responsive changes to the design of the building. They first propose a workflow based on simulations which generate synthetic data to create generative design workflows. This is a three step process where they (1) use Dynamo to implement a generative algorithm which generates many design options (2) Use Revit to create analytical models out of all the Dynamo options (3) connect Dynamo to Autodesk Insight to conduct energy simulations. However, this process is time-consuming and can’t keep up with the fast nature of the iterative design process. Next, they discuss how to use ML models to achieve faster analysis results than simulation models. This is a workflow which essentially frontloads the simulation part of the process. First, they compresses the synthetic data points into a feature vector that represents different variables like height, width, WWR, and HVAC type. Then, they train a model on these x-values, to predict the y-value, which they set as the Energy Use Intensity (EUI). Finally, they can use the model to instantly predict the EUI for new X-values of building forms that haven’t been seen before.
One thing that I was surprised about was how they solved the issue of data scarcity by generating synthetic data and then compressed that data into a feature vector. From my understanding of ML models, the model has to be fed numerical data. For other types of data, such as text, image, or sounds, the conversion to numbers is relatively straightforward (ex. images can be converted to pixel values). I found it innovative that they solved for this issue by encoding the geometry of the model into formats such as voxels, point clouds, and 2D images. However, one uncertainty I have is the generalizability of such models to more complex building shapes. From the presentation, it seems like this solution works great for standard buildings with easily changeable parameters like width and height. But, I’m curious how it could be implemented in buildings with complex curves (like Zaha Hadid or Frank Gehry’s buildings). If I had a deeper understanding of training models with these types of data formats (such as point clouds), I would be interested in trying to create such a model myself.
Before learning visual scripting languages this quarter, I had already taken a few CS classes, and gotten used to programming in Python. There was a bit of friction in adapting to node-based code with inputs and outputs, as I found the logic a bit different from that used in regular code. I did, however, really enjoy that the output was instant i.e. I could see the results of the Grasshopper nodes in the Rhino tab, without needing to ‘run’ the code. The biggest source of friction for me wasn’t time, rather it was the limited amount of Grasshopper tutorials. This led me to find videos on learning Grasshopper and look through blog posts on how to do certain workflows. An AI-version of the workflow could be including a Grasshopper-native LLM, whereby you could prompt an LLM to output the nodes that you’d need (I’ve heard that a similar feature will be released soon on Revit). I could imagine that this LLM would be specifically trained on Grasshopper queries, receiving text prompts as inputs, and a series of nodes as outputs. It’s like vibe-coding but for Grasshopper. Yet, as is the risk with vibe-coding in general, the programmer loses agency over the output and can get lost in ‘sloppy’ code. Something similar could happen here, where a Grasshopper novice could get stuck with a node output that isn’t the most optimal or efficient.
SmartHopper is an open-source project developed to solve the exact need I had in the previous question, an AI assistant within the Grasshopper environment. It was released in July of 2025, so less than a year ago, and works with some of the largest creators of AI companies, such as OpenAI. This kind of tool would have been very useful for debugging scripts within Grasshopper, and getting assistance from a tool that can be customized to every users needs.
Forma Site Design is a new Autodesk cloud-based software that provides instant early-stage site and climate analysis. It can calculate metrics for any given site, such as daylight, sun hours, wind, embodied carbon, and more. I believe this tool is interesting because it removes a big barrier to entry for other, older tools like Climate Consultant or Ladybug. For instance, conducting solar analyses in Ladybug has a higher learning curve, as it requires knowledge of Rhino → Grasshopper → Ladybug. Meanwhile, Forma’s interface makes it easy to run analyses. The climate simulations and calculations rely on machine learning models to predict the output. While I don’t think it would’ve changed how I worked this quarter, it has already been very useful for my architecture studio projects.
Maket is a floor plan generator that takes text as input and creates residential floorplans. You can input requirements such as how many rooms, the shape and square footage. Then, Maket generated the layout and allows you to experiment with different styles and layouts. This tool is interesting because unlike image generators, like Nano Banana, which don’t have the spatial understanding to create floor plans, Maket is trained only on existing floor plans, thus become a highly-specialized ML tool that understands factors like circulation, and room adjacencies. Again, while this isn’t something that I would have used this quarter, I can see it becoming helpful in my future work in generating different layouts of spaces, which tends to be a tedius job.