Module 8 Questions:
Topic: The technical pipeline: Using Generative Design and Machine Learning for Faster Analysis Feedback (AU 2020)
The talk focused on a major problem in the AEC industry: detailed building performance analysis is often too slow and computationally expensive to keep up with the fast pace of conceptual design. The speakers proposed using generative design workflows in Revit and Dynamo to automatically create large synthetic datasets of building forms, then training machine-learning “surrogate models” that can instantly estimate building energy performance without running full simulations every time. They emphasized that early-stage analysis is critical because designers still have flexibility to make impactful decisions while the cost of changes remains relatively low. What I found most interesting was their strategy for solving the lack of architectural training data. Instead of waiting for firms to share real building models, they generated synthetic datasets themselves using parametric workflows and taxonomy-based geometry categories. The part that I would most like to try myself is connecting generative design workflows with performance prediction in an iterative design process. For me, it is important that the workflow can clearly demonstrate the accuracy and reliability of the prediction process. There are many ML models that estimate energy use, seismic response, or embodied carbon during early-stage design exploration, but I would also want transparency in how the model makes predictions so engineers can understand its limitations rather than treating it as a black box.
For me, the biggest limitation of the current parametric design tools is the lack of adaptivity and flexibility in the workflow. During the quarter, I often ran into problems where a node only accepted a very specific type of input and produced a strict type of output. Even when the information was conceptually the same, the workflow would fail because the data type or structure did not exactly match what the node expected. A lot of time was spent debugging connections, converting data types, or reorganizing node structures instead of actually exploring design ideas. This became especially frustrating when building and testing multiple parametric forms because small changes in the workflow could easily break the entire system. I think an AI- or ML-augmented version of the workflow could make the process much smoother and more intuitive. For example, there could be a built-in AI assistant or chat box inside Dynamo or Grasshopper where users describe their goals in natural language, and the software automatically generates the necessary nodes and connections. The AI could also detect errors in the workflow, suggest fixes for incompatible data types, or automatically adapt inputs and outputs between nodes. This would reduce repetitive setup work and allow designers to focus more on evaluating design performance and exploring tradeoffs rather than troubleshooting technical issues. At the same time, I do think some of the friction is valuable because it forces users to understand the logic behind parametric workflows and data structures. Learning how nodes interact and how information flows through the system helped me better understand computational design principles. Because of that, I would not want AI to completely automate the process. Instead, I would prefer AI as a collaborative assistant that speeds up repetitive or frustrating tasks while still allowing designers to understand and control the workflow themselves.
- TestFit is an AI-assisted site planning and feasibility platform that can automatically generate apartment layouts, parking arrangements, and massing studies based on zoning rules and user-defined constraints. Instead of manually drawing every parking stall or testing multiple building configurations one by one, the software rapidly generates optimized site layouts in real time. What makes it interesting is how it combines parametric logic, generative design, and real estate analysis into one workflow. It is not just creating geometry; it also evaluates density, parking efficiency, and development potential simultaneously. During this quarter, this kind of tool would have significantly reduced the repetitive work involved in testing different building forms and site configurations. Instead of manually rebuilding parametric models every time I adjusted constraints, I could have focused more on comparing design tradeoffs and performance outcomes.
- Hypar is a cloud-based generative design platform that allows users to create building layouts and automate design workflows using reusable computational logic. It can generate floor plans, organize spaces, and export directly into BIM workflows such as Revit. The platform is especially interesting because it treats design workflows almost like software modules that can be shared and reused by different teams. What I find most innovative is the idea that architects and engineers can publish and distribute design logic itself, not just final models. This could have changed how I worked during the quarter because many of my Dynamo workflows required repetitive setup and debugging of node connections. A platform like Hypar could potentially automate large portions of those workflows while still keeping the parametric logic transparent and editable. It also aligns closely with my interest in computational structural and architectural design.
- tool is an AI-driven building performance analysis platform focused on sustainability, energy modeling, daylighting, and embodied carbon analysis. The tool uses machine learning and automated simulation workflows to give designers fast feedback during early-stage design. Instead of requiring hours or days of detailed energy modeling, it can quickly estimate performance impacts and compare design alternatives. This is especially interesting because it directly addresses one of the biggest problems in design workflows: simulation speed. During the quarter, I spent a lot of time iterating building forms and evaluating performance metrics manually. A tool like cove.tool could have accelerated that process by giving immediate feedback on how geometry changes affect energy use or environmental performance.