Part 1: Study One Talk
Talk selected: A Practical Use of Machine Learning in the AEC Industry by KLH Engineers, AU 2019
The talk focused on a very practical problem in the AEC industry: converting inconsistent AutoCAD layer names into usable Revit elements. KLH Engineers receives many 2D AutoCAD drawings from outside teams, but the layer naming is not always consistent. Before they could use their own automated tools, someone had to manually clean up the AutoCAD layers and decide what each layer represented, such as walls, furniture, or other building elements. The speakers’ approach was to use machine learning in Python to learn from past manual conversions and predict how new layer names should be categorized. Autodesk describes the case study as using Python to bridge AutoCAD and Revit by converting thousands of possible AutoCAD layer names into usable Revit elements.
What surprised me was that the machine learning use case was not futuristic or abstract. It was not about replacing designers or automatically designing a whole building. It was about reducing a repetitive task that already had a lot of historical data behind it. KLH had accumulated about 38,000 layer-name translations from previous manual work, then used that history to train a model instead of writing every rule manually. This made the talk feel more realistic to me. I also liked that the ML tool depended on the company’s own data and workflow standards. The pushback I would have is that this kind of model is only useful if the company has enough clean past data. If the past data is inconsistent, then the model could just automate bad habits.
I would want to try something similar for repetitive BIM cleanup or Dynamo/Revit troubleshooting. For example, if a firm had many past Revit models, Dynamo graphs, and correction logs, an ML tool could suggest which families, parameters, levels, or categories need to be fixed before analysis. For me to actually trust it in a project, the model would need to show confidence levels and allow human review before making changes. I would not want it to silently change the model, but I would want it to identify likely problems and suggest fixes.
Part 2: Reflect on the Quarter
One moment where I felt the limit of the current tools was when I was working between Revit, Dynamo, and Generative Design. Small setup issues took much longer than expected. For example, mass visibility, mass floors, levels, input and output nodes, dependencies, and custom study setup all had to be correct before I could even evaluate the design alternatives. The actual design question was interesting, such as how height, twist, or floor area changed the performance metrics. However, a lot of my time went into making sure the graph was connected correctly and that Revit and Dynamo were reading the same geometry.
An AI-augmented version of this workflow could act like a “model and graph checker.” Instead of designing the building for me, it could review the Revit model and Dynamo graph and say: your mass floors are not active, this node should be marked as an output, this input is not connected to the study type, or this evaluator is placed before the transaction is complete. It could also explain why a metric is not updating when the design variable changes. I would want that kind of augmentation because it would reduce unnecessary frustration and let me focus more on the design tradeoffs. At the same time, I think some friction was useful because it forced me to understand the relationship between geometry, parameters, and evaluators. I would not want AI to remove the learning process completely, but I would want it to reduce the time spent on technical setup errors.
Part 3: Scout the Frontier
1. Autodesk Forma Building Design
Autodesk Forma Building Design is a cloud-based schematic design and analysis tool that helps architects explore early design options before moving into Revit. Autodesk describes it as a tool for shaping multiple options quickly, testing performance, and then moving the design directly into Revit as a native model. It includes design automation for floor plans and facades, automatic area metrics, sun hours, daylight potential, and carbon analysis. This is interesting because it addresses a problem I felt this quarter: early design iteration is powerful, but setting up a full Revit and Dynamo workflow can be heavy. If Forma had been part of my workflow, I could have tested massing, daylight, sun exposure, and basic performance earlier without building as much setup logic myself. It would not replace Dynamo, but it could make the first round of exploration faster.
2. TestFit Site Planning AI
TestFit’s Site Planning AI helps generate site plans from user-defined requirements such as unit count, unit mix, setbacks, parking count, parking ratio, and other development constraints. The tool then generates a buildable site solution and updates the result when the user changes parameters or site conditions. I find this interesting because it connects design with feasibility. In class, we focused on parametric form, metrics, and tradeoffs, but in a real project the design is also shaped by zoning, parking, density, and financial assumptions. TestFit would have changed my workflow by making the early design variables more realistic. Instead of only testing abstract tower inputs, I could imagine testing options against site constraints and development goals from the beginning.
3. Finch
Finch is an AI and generative design tool for building design that helps teams generate floor plans, test design options, and get real-time feedback. Its product page says it can generate floor plans based on a firm’s plan library and automatically populate units with options that meet code and accessibility requirements. It also describes an AI agent that handles repetitive precision work such as door placement, compliance checks, and consistent updates across linked units. This is interesting because it targets the repetitive interior planning work that often happens after massing. In my quarter, I spent more time on mass form, evaluators, and Dynamo logic, but if the assignment moved deeper into actual floor planning, Finch could have helped convert a massing idea into more usable layouts. It would have made the bridge between conceptual mass and architectural plan much smoother.
Closing Reflection
Overall, this week’s topic made AI in AEC feel more practical to me. The most useful AI tools are not necessarily the ones that claim to design an entire building automatically. The more believable tools are the ones that reduce repetitive translation, setup, checking, and iteration work. After this quarter, I would want AI to support the designer by making parametric workflows easier to set up, easier to debug, and easier to connect to real project constraints. I would still want the human designer to make the final judgment, especially when tradeoffs are involved.