Part 1:
I chose to watch the talk “Using Generative Design and Machine Learning for Faster Analysis Feedback.” The goal is to provide better building stage analysis tools early on that can help customers find higher performance yet low cost designs. They are also looking to use a tool that is less computationally expensive, but also allow quicker updating to follow workflow. The machine learning model they want to use for this does not have enough good data for training, so instead, they use generative design workflows to create synthetic data in Dynamo. The surrogate model is fit to this dataset, and this is validated with the models generated in Dynamo and Generative Workflow again. Something important within the data generation is to ensure that there are no repeated items and overlapping items that will make the machine learning model more biased.
I am surprised that you are able to create your own data to be used for training the machine learning model. More typically, I often see MLMs being trained on real life, existing data to be able to reproduce variations of what they’ve learned, with MLMs of less training making key mistakes that may match what they’ve learned but is not realistic. Hence, I wonder if the dataset generated for training here will have similar issues, or if there is enough breakdown and sorting of generated models before actually feeding it into the MLM.
Their current example looks to be a simple building with simple goal outputs, but it is interesting to see how the EUI generates quickly/live in Revit. I would want to see how this model applies to a more complicated model running multiple different changes and goal variables, and see how fast that generation is prior to implementing a model like this.
Part 2:
Throughout the last few weeks, I mostly saw difficulties with running a model quickly enough, which may be due to how efficient my the workflow I created was. In the beginning, before knowing how to run multiple inputs simultaneously, I would have to go in change the inputs until I got a few results I wanted to report. Even after learning out to generate more outputs, I think the model would take a while to run. Another aspect is that I would typically run into node errors where the inputs I connected are not what the node expects, but the app would not give enough information for me to easily resolve the issue.
Having a ML tool that can help make workflows more efficient and give more detailed information about errors could be helpful, since I think that is what is slowing down how the model is run. This would help me be able to build more complicated workflows and still be able to produce multiple results. However, I do feel like having a tool like this can be a major crutch that would stop users from learning more about the different nodes and tools already available in Dynamo/Grasshopper.
Part 3:
Snaptrude - AI Floor Plan Generator
This tool can get inputs from a spreadsheet that has inconsistent formatting that would typically be hard to read, and uses AI tools to generate a 3D layout with the given information. The model can be updated by just pasting the spreadsheet information, and reload the 3D layout. This makes the process getting information from clients to a computer model faster, and allows a more streamlined workflow where changes can be made very easily and with less human work. This specific tool wouldn’t have changed in the projects I’ve worked on this quarter, but something of the likes could have been helpful. Especially for the modules where we had to generate multiple inputs for processing, it may have been easier to write out some input values and constants in a spreadsheet rather than as blocks of code inside Dynamo/Grasshopper. A tool like this that will pull all my data from a spreadsheet (which is a more common tool across industries) could ease that process.
Pathw.ai - Identifying given steel design and simplifying connection designs
This AI tool analyzes a designed structure (steel beams and columns, and connections, are already designed). Then, it will group similar connection details and generate a simple model of the information fed. Then, the tool will go through preset connection designs, especially from a set of pre-designed connections that may already be used within your firm, and organizes where and how many of each connection is there. The goal is to simplify the information provided to drafters and construction. I think this is an interesting tool as this doesn’t necessarily replace the need for engineers, and instead is used to streamline communication to other stakeholders in the project. Something like this tool could have been helpful in the recent module using generative design, as I was trying to size all the beams and columns in a simplified, preliminary way that can provide results but not require a full set of structural calculations. Something like this tool that can simplify connection, and potentially beam/column sizing, can make my modeling easier and provide faster, visual models if passing this project on to architects.
Sketch To Render - Render 3D drawings and videos from sketches
This tool allows your sketches or AutoCAD/SketchUp models to be easily generated into rendered architectural images and 3D models. It also allows video and walkthrough renderings. This is particularly interesting since it can make architecture/structural design more accessible to people in the industry. It can also provide a more detailed model that can help clients better visualize certain material and facade designs without needing to fully a model the building, saving time for the architects and BIM managers. This tool can be very helpful for some of the past modules I’ve worked on this quarter, especially the ones where I only generate conceptual masses and lofted shapes. I couldn’t actually add any interior or facade designs within Dynamo/Grasshopper. This app, if compatible with sketches from these two softwares, can make my final building mock-ups more realistic and presentable to clients.