Yan Liu

The problem mentioned is that building buildings are responsible for 40% of the global carbon emissions. The traditional detailed energy modeling is too slow for conceptual design so the proposed solution is to use generative design in revit and dynamo to create millions of sample structure and there energy results. This way designers and get a real time estimate on their design because the AI will provide an instant building energy estimates.

The thing that surprising is the huge complexity of all the geometric forms that were created with 5 different climate zones. I would have never thought to include that if I was making this design. My pushback would be instead of using generative AI to create these models and generating the data why not use real data and capture real models of current buildings and projects and use those scores to predict and lessen carbon foot print. I know its a higher upfront cost.

I would want to try this generative ai only if it was integrated with cost. Maybe have a bunch of designs and their cost with carbon footprint ratio so I can decided at the same time the cheapest option.

The problem at hand here is the massive shortage of AEC skilled labor. It is projected to reach 750,000 vacancies by 2026. He proposes to switch to using algorithms to solve complex task like test fitting directly in Revit and have some sort of automation of fabrication documents to bridge architectural and manufacturing.

Something exciting about this process is that they are approaching this problem like a car manufacturing process. A very systematic solution to a complex muti variable industry. I would push back on the assertion that design "cannot simply be a matter of intuition" While data optimization is powerful for complex problems, human intuition is still vital for navigating subjective aesthetics that algorithms cannot yet quantify.

I would want to trial the Design to Fabrication workflow to automate the generation of 3D views and sheets for unique parts. For this to work, a project would need integration between Revit, And Rhino, and other data sources to work efficiently.

There are inconsistent AutoCAD layer names from external architects. There is a manual cleaning that needs to take place to standardize these layers. The solution is to use ML to look at the layer itself and predicts the category it belongs to.

A major surprise was that the model eventually became more accurate than the human designers! I can’t think of a pushback for this method it seems like an perfect way to use ML/AI to get rid of mundane work.

I want to try their confidence based user experience, where the machine color-codes its decisions so a human only needs to verify the ones the machine is unsure about.

One moment that clearly stood out to me was building the shapes for the structures. Either it being a square, circle, star, or any other shape it always started out with the same 0,0,0 point and copying it and transforming it and connecting those points. Within dynamo there is kind of an AI took which tries to guess what node comes after but I feel like using an AI or ML version, it would know what I was trying to make and suggest it through a preview and I would have to hit yes or no to train it. If yes then all those nodes are placed for you and if no it will see what I did build and learn that.

Some other friction points would be invalid inputs to a node. I would have to use AI to help me and the solutions would be something super random I wouldn’t think of, like flatten the list, check the cross product, check the leveling, etc. Maybe AI/ML could help with this aswell, if an errors occurs from an input run and check all those niche check boxes and show it works to the user and let them know the adjustment made. This process should be automated so its easier to work with loads of data and information.

Some cool technologies I found:

Reconstruct - AI Progress Monitoring

Reconstruct turns job site footage ,shot on a phone, 360 cam, or drone into a live digital twin you can compare against the BIM model and schedule. The tech uses deep learning on both images and 3D point clouds to identify what's actually been installed and flag deviations from design. And you can rewind the site to any earlier date. Traditional progress monitoring is a super manual exercise because someone needs to walk into the site, take photos that go into a report, and that report goes into a meeting three days later. Reconstruct compresses that into near real-time deviation detection. I had no need for this tool this quarter but in the field this would be pretty helpful.

Speckle - Open-Source BIM Data Layer

Speckle is an open-source platform that lets geometry and project data move freely between Revit, Rhino, Grasshopper, Dynamo, Python, and a dozen other tools that normally don't talk to each other. It started as a PhD research project at UCL and now its a too for many firms. Firms use it to route their BIM data automatically into their estimation system and use AI to give them estimates. Its a cool platform that allows the mix of different types of data for the use of training or giving to AI without the data being stuck in only revit or only rhino etc etc. This would of been helpful to be this quarter because sometimes I wanted to try something in Rhino but I already started my homework in Revit and I didn’t want to rebuild everything. Would of saved me time to easily transistion between the two.

Notebook LLM - Googles AI research tool

NotebookLM is Google's AI research assistant that lets you upload your own documents PDFs, slides, notes, etc and then chat with them directly. Instead of a general chatbot pulling from the internet, it grounds every answer in your specific sources and cites exactly where it got the information. It has cool features like iAudio Overviews, which turns your documents into a two-host podcast-style conversation summarizing the material I’ve been using it for my FE prep. I inputed so many NCEES reference handbook plus notes had it generate practice questions, and explain concepts.