Module 8 Questions:
Part 1 - Study one talk
Option C: A Practical Use of Machine Learning in the AEC Industry
After working as a civil engineer for three years, I can attest to how time-consuming and frustrating the AutoCAD-to-Revit layer conversion process is. Even when working meticulously, you finish unsatisfied, as architects and other external team members hand you endless ambiguous layer names with no consistent convention. As such, this talk piqued my interest.
The problem the KLH engineers were trying to solve was how best to automate the translation of thousands of inconsistent, external layer names into standardized Revit elements. Their team chose machine learning over rule-based programming because the possible layer names is effectively limitless, so hard-coded rules could never suffice. The ML model was able to pick up on underlying patterns in how names map to elements, including the weird edge cases most engineers would not anticipate.
It was surprising to learn that the team plateaued at an 80% validation rate and seemed comfortable stopping there. Their explanation was that the model was trained on human-labeled historical data, and since humans themselves perform below 80%, the model inherited these bad patterns. That logic may hold, but I would push back on accepting it for a tool being marketed as finished. Cleaning even a subset of the training data to push validation toward 90% seems worth the effort.
The advice I valued most was not to reinvent the wheel. Use publicly available, free resources, including existing models on GitHub, rather than building from scratch. It’s certainly important to understand the basics of how a model works and how it reads your data, then feed it your own. I’d want to try this myself, and I’ve already sent the talk to a few former colleagues.
Bonus
I decided to watch the other two lectures as well. Please see below for my brief take on both of the talks:
Option A - Using Generative Design and Machine Learning for Faster Analysis Feedback
- The talk starts with hammering home the statistics: buildings account for 40% of global carbon emissions, and we need to cut this by 50% by 2030. While most designers agree that decisions in the early phases of a project have the highest impact on performance, there is a disconnect in that getting energy feedback on designs is too slow to be useful at that stage. Thus, folks at Autodesk decided to train ML models to predict building energy in milliseconds instead of waiting on full simulations.
- What’s interesting is that they admitted that real AEC data is too scarce and inconsistent to train ML models, so they were forced to generate synthetic data instead. Using generative design in Revit & Dynamo, the speakers created almost 1,000 building geometries across 6 shape types. From there, they ran 250 parameter variations per form across 5 climate zones, which allowed them to achieve about a million labeled data samples. Their results actually exceeded my expectations as they had an approx. 1.9% prediction error in their model, which is incredibly low. This technology allows architects and engineers to have instant estimates on building energy during the earliest conceptual design stages.
Option B - The Future of BIM is NOT BIM, And It’s Coming Faster Than You Think
- This was a really neat talk, especially considering it was presented in 2019, now seven years ago. What stood out most was how broad the discussion was, covering a wide range of emerging AI applications across the AEC industry. Bill Allen highlighted several technologies that are now becoming increasingly relevant, particularly the use of LiDAR scanners and VR headsets within construction.
- One of the strongest themes throughout the presentation was the construction industry’s resistance to change. The industry has long operated under the mindset of “if it ain’t broke, don’t fix it,” which often slows the adoption of new technologies. Allen emphasized that this mindset must begin to shift if the industry hopes to address growing global and environmental challenges.
- His point about the housing crisis was especially compelling. The United States, along with many other developed nations, faces a major shortage of housing, and the current pace of construction is simply not sufficient to meet demand. Integrating technologies such as AI-driven workflows, digital modeling, LiDAR scanning, and virtual reality could improve efficiency, coordination, and speed throughout the design and construction process.
Part 2 - Reflect on the quarter
One friction point I had this quarter was the optimization of the Dubai tower, where I was forced to use a brute-force Colibri sweep since Octopus (the multi-objective solver) only works on Windows. I conducted 50 cases across twist and side-count, exported to CSV, then manually filtered by GFA and ranked with a weighted composite score in google sheets. An evolutionary optimizer would have searched the space intelligently instead of forcing me to evaluate every combination.
A machine learning model that was trained on a subset of evaluated designs could predict solar and view scores nearly instantly, which would have made a real optimization loop feasible even on a mac. Would I want that? Eh, yes and no.
For the individual sweeping, yes, the brute-force grid and the manual sheets work was not valuable. But, I was able to build some intuition for how twist and side count traded off against each other by looking at the results myself. In addition, it forced me to create a weighted score which added some additional judgement that I’m thankful for having. So, I would have automated the calculations/exporting but not the decision/analysis parts.
Part 3 - Scout the frontier
- Testfit: https://www.testfit.io/
- Testfit allows developers and civil engineers rapidly iterate through fully resolved site plans given site constraints and local code parameters. As a civil engineer, one of these basic layouts took me one to two hours to draft as an exhibit, only for the client to then cycle through a dozen variations, each requiring a near-total redo. I always assumed in the back of my head that this could be automated, and this assignment pushed me to actually find the product doing it. While most of this is built using a proprietary parametric algorithm, the generative design feature lets the system search configurations on its own against goals like yield on cost rather than only reacting to manual edits. In addition, it has an earthwork optimization that minimizes cut-and-fill volume directly, which was a really tedious task to figure out as a civil engineer.
- This product would not necessarily change how I worked this quarter since we did little with site layout, but it certainly would have impacted my time at Bohler working as a civil engineer for three years.
- Asterisk (Thornton Tomasetti):
- I will be working at TT next fall, so this product is particularly interesting for me. This product uses custom ML and generative AI models to help their engineers rapidly explore design options efficiently. Users can apply specific design parameters to generate a wide variety of structural concepts with rapid member sizing, quantity takeoffs and embodied-carbon calculations. The design includes floor and column placement, embodied-carbon estimates, total area, cost per square foot, and weight. Thus, TT can go to their clients armed with key performance indicators at a fraction of the cost during the early stages of a project. In my opinion, this product would give them a huge edge over competitors until others catch up.
- This definitely would have changed how I worked this quarter as it would have been excellent for the Dubai project.
- Build Check.AI:
- This one is actually a Stanford-founded startup! The idea is that the AI program reviews 2D construction drawings across all disciplines (architect, structural, civil, MEP, landscape, etc.) at once. It then runs hundreds of checks to detect errors, omissions, and coordination issues that humans typically miss before they snowball into expensive field issues or RFIs. It basically polices the drawings for costly mistakes that would otherwise surface downstream.
- Similar to Testfit, this product wouldn’t have helped much for this class since we didn’t have to coordinate with other disciplines; however, it would have been very helpful during my career as a civil engineer and will likely be used in the future as a structural engineer.
