Part 1
For Part 1, I chose the AU 2020 talk about using ML for faster analysis. The main problem the speakers are trying to solve is that running building performance simulations, specifically for Energy Use Intensity (EUI) and how its way too slow during early conceptual design. You can't iterate quickly if you have to wait hours for a simulation. To fix this, they propose using a machine learning surrogate model to predict the results instantly. Because the AEC industry doesn't have the massive 3D datasets needed to train these models, their approach is to manufacture a synthetic dataset. They used Generative Design in Revit, Dynamo, and Insight to generate and simulate thousands of building massings, ultimately creating about a million labeled data points to train their ML model.
What genuinely surprised me was how they treated Generative Design. Instead of using the solvers to find the "optimum" design, they essentially just used it as a giant data factory to feed their algorithm. However, I’d push back a bit on the real-world reliability of this. They trained their model on idealized, parametric massings. In reality, site constraints, structural irregularities, and constructability issues make projects incredibly messy. I’m skeptical about whether an ML model trained in a pristine synthetic environment can accurately predict performance for a highly irregular, real-world building.
To try this on a real project, maybe applying the same synthetic data pipeline to something like preliminary structural sizing, I would need a really transparent way to validate the ML's predictions.
Part 2
Looking back at the parametric modeling assignments over the last eight weeks, my biggest friction points often had less to do with high-level design logic and more to do with just getting Dynamo to cooperate with Revit. I spent a surprising amount of time troubleshooting things that felt like they should be straightforward, like extracting the correct curve lengths, getting surface panel heights to adjust predictably, and keeping my code blocks organized with comments. Even just managing the links and navigating back and forth between the Revit model and the Dynamo script felt clunky and slowed down my workflow.
If I could have an ML tool here, I'd want a "smart assistant" or auto-troubleshooter right inside the Dynamo workspace. Instead of me hunting through forums trying to figure out why a node is outputting null values for my panel heights, the AI could look at my script, flag the broken logic (like a list lacing issue), and suggest the correct node structure.
I would absolutely want this kind of augmentation. The friction of fighting the software interface or troubleshooting list levels doesn't actually teach me anything meaningful about design or parametric geometry; it just eats up time. If an AI could handle the translation between my intent and Dynamo's specific node requirements, I could spend that time actually evaluating the design tradeoffs.
Part 3
1. Hypar
- What it does: Hypar is a cloud-based generative design platform that allows users to write algorithms to automatically generate building systems. More recently, they've integrated text-to-BIM capabilities where you can literally type a description of a building, and it will generate the 3D geometry and underlying data in the browser.
- Why it's interesting: It takes the logic of visual programming tools like Dynamo and moves it entirely to the cloud. Instead of passing around massive, heavy scripts that might break depending on your local software version, different disciplines can combine their generative algorithms into one lightweight, federated model on the web.
- Impact on the quarter: Instead of struggling to keep my local Revit and Dynamo instances perfectly synced without crashing, I could have used Hypar to run and evaluate parametric tradeoffs. It abstracts away a lot of the visual programming friction (like managing list structures) and just delivers the generated geometry, which would have let me test way more variations in the same amount of time.
2. Veras (by EvolveLAB)
- What it does: Veras is an AI-powered visualization plugin that works directly inside Revit and Rhino. It uses text prompts to generate high-quality, conceptual renders and design variations based entirely on the 3D geometry sitting in your active viewport.
- Why it's interesting: It solves the massive time-sink of conceptual rendering. You don't have to spend hours applying materials, tweaking lighting, or modeling site context just to see if a structural form look right, the AI interprets your generic massing and visualizes it instead.
- Impact on the quarter: When I was flexing different forms and evaluating tradeoffs this quarter, my models were mostly just gray boxes and wireframes. Being able to run a quick text prompt to see how those different parametric iterations would actually look with a glass facade or concrete detailing would have made it much easier to evaluate the qualitative, architectural aspects of the design alongside just numbers.
3. Cove.tool (ML Optimization)
- What it does: Cove.tool is a building performance analysis platform that uses machine learning to automate energy, daylight, and cost simulations. It runs thousands of iterations in the cloud to find the most optimized design variables (like window-to-wall ratios or insulation thickness).
- Why it's interesting: It essentially operationalizes the exact theory described in the AU 2020 talk from Part 1. It takes the incredibly painful, slow process of setting up detailed analytical models and uses ML to give you near-instant performance feedback during the early design phase.
- Impact on the quarter: Rather than manually rigging up analytical spaces and waiting for traditional simulation engines to parse the results, I could have plugged my parametric geometry directly into their ML engine. It would have allowed me to plot complex tradeoffs (like energy performance vs. upfront carbon) across thousands of generated options, rather than just the handful of iterations I had the patience to manually test this quarter.