Module 8 Questions
I chose to watch A Practical Use of Machine Learning in the AEC Industry, covering a concrete case study of using ML applications within existing design platforms. The speakers were trying to aid designers at KLH in transitioning layers received in AutoCAD to those that could be named and applied in REVIT. Typically a rather labor and time intensive process, the ML was designed to help cut down both on the time intensity of tasks for designers and to generally bridge the existing gap between subcontractor workflows and varying applications used. Specifically, the ML in this case served as a layer name translator, that was trained on past name conversion data stored in csv files, and looked at letter pairing within AutoCAD layer names to construct layers as a point in a very large multidimensional space, then correlating to a new name within a standardized REVIT layer naming scheme. When applied to projects, the ML output both the best approximation of translated name as well as confidence in translation, allowing for designers to concentrate effort and time where necessary.
One thing in the talk that genuinely surprised me was when they were discussing the initial implementation of the ML tool in the company. Initially, they were hesitant to release the tool because it was only measuring at a 80% accuracy rate. However after investigation, it was revealed that this was largely due to human error present in the training data, and the ML was actually able to more accurately predict proper layer names because it assessed all possible options rather than maybe selecting the first seemingly right alternative on the dropdown. I thought this was interesting because it shows how susceptible current design is to human error and error propagation, and the potential of ML and AI to not only decrease time spent on more monotonous tasks and streamline design, but also increase the accuracy as well as efficiency. I think it's important to note though that it was so important to color the ML named layers as different and distinguishable from the human identified ones, because even though the tool may sometimes be more accurate, it relies on human data training and ultimately is no substitute for the experience and decision making abilities of skilled and seasoned designers.
I think in order for me to use tools like these on a project, it would have to be very obvious what outputs were ML or AI generated. I think there is value in using tools such as these to accelerate design, but I think blind confidence in generated solutions can cause several issues. Even a rule of thumb or basic check of generated results would be very important in verifying the case by case solutions presented by ML. Especially in design and engineering where no decision exists in a vacuum, using AI generated solutions as inputs to later aspects of design could be detrimental if an issue arises either in training or application.
For me, the primary limitations of the current tools centered around troubleshooting capabilities. When logic within the interfaces wasn’t working, it was easy to discern at which node that may be occurring using watch nodes in Dynamo or looking at the coloration of nodes in Grasshopper, however there was little input on what may specifically be going wrong. I remember in Module 6 I used one of the pre-designed custom nodes to try to estimate development cost via floor area and height along the building, however the logic for mass floor IDing wasn’t working, and it took several hours of research and troubleshooting to identify another solution to achieve a similar metric. Later in the generative design assignment, Module 7, Dynamo provided a little more feedback when my study initially wouldn’t translate to “Generative Design,” saying that the study needed to be rerun with corrected inputs and outputs, which was much more helpful in identifying and resolving the issue. A lot of times when there was a severe problem, a node would also output either null results or all zero. I imagine this would make it relatively easy to integrate some form of an AI addition that could identify these bottlenecks, and possibly provide input on how to fix the current logic to at least align with the node input and output needs if they aren’t in congruence with the current workflow. When necessary, it was possible to ask existing AI tools such as Claude to provide suggestions on issues in the node networks and possible solutions or suggestions for alternate nodes, which also makes me think it would be relatively easy to implement an AI assistant into the existing interfaces.
Tool 1 - Rayon (2021) : Rayon is an AI-based design application that enables architects to generate floor plans and interior layouts quickly. The interface works between BIM and CAD applications, and accepts user input via text requests or alterations to existing generations. Options to start from scratch or input in progress design elements are both available. I think this is an interesting application because it may allow for more efficient floor plans and unit design going forward. I wasn’t able to fully glean the capabilities of the application, but if it is able to quickly iterate through various design layouts to maximize building characteristics such as daylight penetration via extended window area or centralization of wet core plumbing features, that may be really beneficial increasing not only design speed but general occupant satisfaction and building efficiency. I don’t think this would’ve been applicable to my work in this class this quarter, but may have been helpful in a different project I was working on that required the use of unit configuration layouts in an affordable housing project.
Tool 2 - MELT Plan: MELT is a company with multiple subsidiaries, including MELT Code and MELT Takeoff. Both are AI-based applications that help either those in the design or construction industry. Code allows for easy queries concerning building codes or general building questions in regards to guideline or standards compliance. Takeoff allows for input of current design plans and estimates material, construction, and labor costs associated with creation. Both of the applications help accelerate design or estimation timelines, and makes synthesis of existing resources easier and more easily usable. This could have been beneficial in doing the cost estimate of the building design to incorporate more information of construction materials and labor, though the design generated in the modules may not have been fleshed out enough to apply these metrics.
Tool 3 - Helix: Plugin to convert AutoCAD or Sketchup files to REVIT components, allowing for various models to sync together despite original design platform and to streamline future optimization processes. This would be very useful in application when different subcontractors may submit plans in varying design platforms. I think this would be really helpful if there were a lot of varying models going into one design, or if optimization wanted to be done using Dynamo because the forms could be converted to be REVIT compatible.