Module 8 Questions:
Part 1:
In the “Using Generative Design and Machine Learning for Faster Analysis Feedback” video and handout, the speakers are trying to solve the problem of traditional based building performance analysis. They state that it is slow, disruptive, and computationally expensive, which is an issue when thinking about the need for rapid iterations in early design. Oftentimes, engineers and architects need fast feedback in order to make decisions, but the reality is that creating performance models take a lot of time, and are expensive, this means that performance doesn’t always happen as early in the process as it should.
They propose the use of machine learning and generative design to get faster feedback. First, they talk about using Dynamo and Revit to create generative design scripts that are able to instantly create a lot of different building shapes and options. They then propose the use of Autodesk Insight to create a large synthetic dataset that trains a machine learning model to predict the building performance. They are basically saying that they don’t need to run full analysis models every time.
One thing that genuinely surprised me was the idea of using generative design not to find the best building option, but to create training data for machine learning. In this class we have spent a lot of time finding the “optimal” solution using these visual programming tools, so it was interesting to see that that wasn’t necessarily the objective of the talk. Its interesting because it changes the way we think about generative design tools, instead of being just for design exploration they can be used as infrastructure for training AI. I would push back on how accurate the models are if what they are being asked isn’t directly on the data they are trained on. I generally think AI hallucinates a lot, and having machine learning models predict building performance could be dangerous if the person interpreting the results doesn’t have a good understanding of what optimal results should look like.
I would want to try a small scale of the proposed workflow in a structural engineering context. Since we’ve learned to use Dynamo, I would want to create multiple structural layouts and train a predictive model for things like embodied carbon, drift, materials. In order to trust it I would need to know that it has enough examples to make sure it isn’t just designing for a certain building type or missing key structural components. I would also want it to integrate with Revit, so once the optimal design was selected it could push the data into Revit, so that I don’t have to spend the extra time modeling.
Creativity Points
Watched the other videos and created bullet point list answering the questions:
The Future of BIM is NOT BIM....And It’s Coming Faster Than You think (The Sequel)
- Problem: The AEC industry is too dependent on manual workflows. People aren’t taking advantage of the capabilities of BIM
- Approach: Using computational design, automation, machine learning and robotics instead of manually modeling everything
- Genuinely surprised me: I was surprised how much they emphasized that the future is not BIM, because I feel like that is all that is talked about in industry right now
- What I’d want to try: I would want to create a computational design workflow in dynamo that gets rid of some of the repetitive modeling tasks in Revit or ETABS, but would need to have it integrate well and actually model the correct elements and not invent data.
A Practical Use of Machine Learning in The AEC Industry
- Problem: Engineering projects and the AEC industry create a lot of data for each project, but firms do not use that data effectively to improve their processes.
- Approach: They propose training machine learning models to identify issues, project risks and automate repetitive tasks.
- Genuinely Surprised me: The speakers showed how accessible the machine learning models can be. I usually think about these things as really advanced and not straightforward to use (because why else would people not be using them).
- What I’d want to try: I would want to use machine learning and Dynamo to automate some of the repetitive tasks and flag inconsistencies.
Part 2:
One of my biggest limitations this quarter was my computing power. A lot of the models that we created in this class were relatively simple, but as soon as I tried to add any geometric complexity or have more iterations, it would take a long time to run or my computer would crash. I spent more time waiting for the models to run than actually focusing on the results. This was really frustrating, especially when I was trying to add multiple parameters. It made me want to create the simplest models I could, so that I could avoid the long wait times and computer crashes.
I could see AI or machine learning helping with this by predicting results instead of having to fully run the model every time. Instead of having a lot of iterations and tests, the AI model could learn from previous tests and help narrow down which ones are worth running. That would make the workflow faster, and it would have allowed me to spend more time thinking about what I was modeling and the results, instead of spending most of my time frustrated. For the purpose of this class I do understand the value of running the iterations yourself, because it allows us to see how the changes in parameters affect everything else. So, in real life I would definitely want that augmentation, and to some degree it would be useful in this class, but I do think that having it completely automated would defeat the purpose.
Part 3:
The first tool I found interesting is one developed by Finch 3D. It is an AI assisted design tool that generates floor plans and building layouts based on parameters given by architects and engineers. It simplifies the process of creating plans, and removes a lot of the manual work. It was interesting to me because it shifts the way we think about making floor plans, they become more interactive and more iterations are able to be seen versus if you were hand drawing them. I think it would have changed the way I worked in some of my other classes, where I had to develop floor plans. In the case of this class I don’t think it would’ve been as relevant because it is a different scope.
The second one I found is called Hypar, which is a platform that uses computational design, automation and AI for building systems. It is interesting because it feels less focused on modeling individual objects and more on defining the relationships between the systems. I think using the logic that hypar has could’ve been helpful in my workflows, where I could have spent more time thinking about having adaptable workflows that would react to the parameters that needed to change, instead of rebuilding. But again, using it entirely would defeat the purpose of learning dynamo and understanding how to make the workflows.
The third tool I found was Carbon Designer 3D by One Click LCA. It uses AI to estimate embodied carbon early in the design, allowing designers to understand the impact even before the full BIM model is created. It is interesting because it takes some of the tediousness out of calculating embodied carbon, and it allows the designers to know the impact really early in the process. This would have been helpful, especially in the assignments where we were able to pick our optimization parameters. I wanted to optimize for embodied carbon but wasn’t sure exactly how to approach it or if the results would be meaningful. Using this tool could have helped me measure the embodied carbon and have metrics that I could compare it to.
Creativity Points:
Another one that I used in another one of my classes was Buro Happolds BHoM pluggin to grasshopper. These tools allowed me to create a grasshopper script that was able to extract a BIM model from Revit and pull it into ETABS. This was very interesting because saved a significant amount of time in the modeling process and it removed the fear of data loss between models.