Module 8 Questions:
I am very interested in all three topics since each talk give a different perspective on how AI and ML may change the built environment, so I decided to review all of them and I am eagerly hoping this can be considered for extra credit.
Option A - Using Generative Design and Machine Learning for Faster Analysis Feedback
This is still my favorite topic, as this talk focuses on how machine learning can make early building performance feedback faster align with my interest and future career. The speakers explain that running energy analysis for many design alternatives can be time-consuming and computationally expensive, especially during conceptual design. Their approach is to use Generative Design in Revit/Dynamo and Insight to generate synthetic datasets, then train a machine learning model to predict energy analysis results more quickly. What surprised me is that the difficult part is not only training the model, but creating a diverse and meaningful dataset first. I used to think ML was mostly about the algorithm, but this talk made me realize that the quality of design parameters and data structure is just as important. I would want to try this for early-stage energy or solar prediction, especially because in my own tower study I had to compare many alternatives using height, rotation, surface area, cost, and solar potential.
Option B - The Future of BIM Is NOT BIM, And It's Coming Faster Than You Think - The Sequel
To be fair, I watched this talk at high speed and also read through the presentation. This talk looks at the larger industry shift from BIM as a manual modeling process toward a more automated, data-driven, and optimization-based workflow. The main idea is that the future may not be about creating one detailed model by hand, but about defining rules, constraints, relationships, and performance goals so computers can generate and evaluate many possible solutions. What I learned from this talk is that BIM is becoming less about “modeling the object” and more about managing information, logic, and decision-making across the project.
One important takeaway for me is that generative design is only useful if the design goals are clearly defined. The tool can create many options, but it does not automatically know which option is meaningful unless the designer sets the right constraints and evaluation metrics. This connects directly to what we did this quarter, especially when we had to choose metrics like cost, surface-to-volume ratio, floor area, and solar potential. I agree with the direction of the talk, but I would also push back on the idea that optimization can fully replace design judgment. A building is not only a result of measurable performance; it also involves context, user experience, aesthetics, constructability, and values that may be difficult to encode into the model. I would want to try this workflow for early massing studies, where AI or generative tools can quickly test options, but I would still want the designer to stay responsible for deciding what “better” actually means.
Option C - A Practical Use of Machine Learning in the AEC Industry
This talk was the most concrete and practical example because KLH Engineers used machine learning to solve a very specific workflow problem: converting inconsistent AutoCAD layer names into proper Revit categories or elements. Instead of presenting ML as a futuristic tool that generates an entire building, they used it to automate a repetitive task that takes a lot of manual time in real AEC practice. Their approach was based on using historical project data, where the model learns from previous layer naming patterns and predicts how new layers should be classified.
What I learned from this talk is that machine learning in AEC does not always need to start with a very ambitious design problem. Sometimes the most useful application is cleaning, translating, or organizing messy project information. This was interesting to me because a lot of friction I experienced this quarter was not only in creating geometry, but in making data readable and usable such as list structure in Dynamo, naming outputs, exporting to Excel, and making sure the model information connects correctly to evaluation metrics. I also liked that this example shows ML as a support tool, not a replacement for engineers or designers. The model can speed up classification, but human review is still needed because wrong classification can create downstream errors in the Revit model. I would want to try a similar approach for organizing BIM elements, classifying model data, or cleaning exported analysis results, especially if a firm already has enough past project examples to train from.
One of my biggest struggles this quarter was that I did not always understand the concept of the tools at first. I was more used to BIM as a way to draw or model geometry directly, like in Revit where I can see the object I am creating. But in Dynamo and Grasshopper, the process is very different. Instead of drawing the geometry, I had to define relationships, data flow, lists, parameters, and logic first, and then the geometry came after. This was honestly confusing for me, especially when the model did not work and the error was only “null” or something related to list structure. A lot of time was spent not on design decisions, but on understanding why the software did not read my geometry, why the image mapping was flipped, or why the solar analysis could not recognize the building form.
An AI-augmented version of this workflow would be really helpful if it could explain the logic behind the node, not only fix the error. For example, it could tell me: “This node needs a surface, but you connected a curve,” or “your Revit mass is not being read as a building form, so the solar analysis returns null.” That kind of help would make the learning process much faster. AI augmentation in this workflow is really helpful and I pro with that, but not in a way that removes the learning process completely. I would still want to understand the concept myself, because I think part of the struggle is also the point of the class. It forced me to think differently about modeling, from drawing objects to building a system. But I think AI is a very useful as a guide, from giving step by step to produce the stage that we want and solving the problem, especially about the hidden technical logic of the tool.
BuildCheck AI is an AI-powered drawing review platform that checks construction drawings for errors, omissions, and coordination conflicts before they become city comments, RFIs, change orders, or field problems. In my own words, it is like an AI reviewer for construction documents: it reads drawings at scale and flags issues that may be missed during manual coordination. What makes it interesting is that it applies AI to a very practical AEC problem, not just form generation. BuildCheck uses custom AI models and computer vision to interpret layouts, annotations, and plan relationships, which helps catch drawing issues earlier in the process. This would have helped me this quarter in a different way: not by creating geometry, but by checking whether my outputs made sense. For example, when my Dynamo or Revit workflow produced null values, wrong surfaces, or confusing exported tables, an AI review layer like BuildCheck could have helped identify inconsistencies earlier before I spent too much time troubleshooting the wrong part of the workflow.
Finch is an AI-powered generative design tool that helps architects generate and refine floor plans quickly. What makes it interesting is that it focuses on the early design stage, where there are many layout possibilities but not enough time to test all of them manually. This relates to my work because I often spent time setting up parametric logic before I could even evaluate the design. Finch would not replace the parametric work, but it could help generate stronger starting options before moving into Revit, Rhino, Dynamo, or Grasshopper for deeper analysis.
Innovative AI I want to develop: “DecarbRx: An AI Decision Platform for Low-Carbon and Energy-efficient Building Design”
The reason why I want to study at Stanford is because I want to develop an AI software that I recently named it, DecarbRx. It is an AI-powered sustainability diagnosis platform for low-carbon and energy-efficient building design. In simple words, it works like a “doctor” for buildings: users upload a BIM model and detailed Bill of Quantities, then the platform checks where the biggest embodied carbon and operational energy problems come from. What makes it interesting is that it does not only calculate carbon or energy performance, but also explains what caused the problem and what the team should do next. This is important because in real projects, architects, structural engineers, and MEP engineers often analyze their own parts separately, so the team may know that the building has a carbon problem but not know which action should come first, how much it will cost, or how much design change it will require. DecarbRx tries to close this gap by ranking recommendations based on carbon savings, energy savings, cost, feasibility, local material availability, and level of design disruption.
The idea will adopt “diagnosis and prescription” workflow. Many tools can show performance numbers, but they do not always translate those numbers into practical next steps. DecarbRx would help answer questions like: should I reduce concrete volume, change façade materials, improve HVAC efficiency, add shading, electrify the system, or use renewable energy? It is also interesting because it can support both new construction and retrofit projects, especially through a Minimum-Change Mode that prioritizes strategies that improve performance without changing the design too much. If I had this tool during this quarter, it would have helped me connect my parametric design alternatives with clearer carbon, energy, and cost recommendations instead of only comparing geometry metrics. It would not replace my design decision, but it would help me understand which option is more worth pursuing and why. Happy to discuss more if you want to know :)