Module 8 Questions:
I chose the KLH Engineers talk on using machine learning to convert AutoCAD layer information into Revit-ready elements. The problem they address is the translation of inconsistent CAD layer naming conventions into standardized BIM data. Because different consultants organize and label information differently, converting legacy documentation into a usable Revit model often requires significant manual classification and interpretation. Their approach was to train a machine-learning model on historical layer translations completed by their team. Using past decisions as training data, the model predicts the likely phase and category for new layer names. I found this approach convincing because the task is narrowly defined: the input is a layer name and the output is a classification that can be reviewed quickly. Rather than attempting to automate design, it automates a repetitive data-management task. One aspect I would push back on is the quality of the training data. Historical project information may contain inconsistent decisions or office-specific habits, which the model could learn and reproduce. If I were to try something similar, I would use a human-in-the-loop workflow where users review predictions, see confidence levels, and correct mistakes. The value of the tool would be reducing repetitive interpretation while keeping professional judgment in the process.
One of the main friction points this quarter was working through legacy nodes, custom nodes, and visual graph organization in Dynamo. The logic of the workflow often made sense, but building it visually was slower than writing code. Node searching, wire management, package compatibility, and debugging list structures all added friction. I also found it limiting that comparing or reusing pieces across graphs was less fluid than in programming environments where multiple scripts can be open side by side. An AI-assisted Dynamo workflow would be useful if it acted as a graph assistant rather than a replacement for understanding the process. It could identify deprecated nodes, suggest alternatives, explain list errors, or generate a draft graph from a written description of the workflow. That would reduce time spent on interface mechanics while preserving the need to understand how data moves through the graph. The most valuable use of AI would be removing repetitive setup and debugging tasks without hiding the underlying logic.
Multi-Agent LLM Structural Analysis Research: This research uses large language models to translate structural problems into a common intermediate format and generate executable scripts for ETABS, SAP2000, and OpenSees. The value is reducing re-modeling effort between platforms and improving interoperability across structural analysis software. It connects directly to the course theme of translating design intent into software-specific workflows.
Foundaxis: This AI-assisted foundation design platform combines foundation layout generation, FEM analysis, geotechnical checks, reinforcement design, and reporting. It is interesting because it automates iterative foundation design tasks while accounting for engineering constraints such as loads, soil conditions, reinforcement, and material quantities. It demonstrates how AI-assisted workflows can support practical structural design decisions rather than only geometric modeling.
Hilti PROFIS Engineering Smart Design: This computational optimization tool evaluates anchor layouts and baseplate connection variables to reduce manual iteration in post-installed anchor design. While not pure machine learning, it targets a repetitive and highly constrained engineering problem where loads, edge distances, embedment, and constructability must all be checked. It reflects a broader trend toward automating checkable engineering tasks while keeping the engineer responsible for final review.