Use machine learning to leverage architects' previous work to enhance future work.
Internship @ Autodesk
Jun- Aug 2019
This project aims to use machine learning to leverage the richness of past data to inform future work for Autodesk Revit architect customers. The ultimate goal is to allow data to be reused and increase efficiency in iterative workflows.
I followed a double diamond design process to define the problem, ideate, and devise a solution. Specifically, I focused on discovering and defining future opportunities. I also collaborated with a Machine Learning Engineer Intern to design and develop a prototype based on what we could achieve from the technical side within limited time and data.
WHAT: Efficiency of referring to previous projects
Architects may find similar projects they did before for reference. How might we make this workflow more efficient?
WHY: Repetitive and tedious tasks for architects
Most architects spent most of their time in the design development phase where they design and draw detail floor plans. However, floor plan drafting is not always innovative. Some types of rooms or even projects tend to be repetitive and architects would refer to similar past projects.
WHO: Experienced architects and entry-level architects
Among all levels of architects, I focused on those who are typically involved in the design development phase. Interns and entry-level architects are those who tend to work a lot on more tedious tasks. And more experienced and senior architects also cannot avoid repetitive work. The opportunity of improving their work efficiency is huge.
I conducted user research with customers and specialists to understand architects' experience of referring to previous projects. Specifically, to get insight into when architects find themselves revisiting their past works and the reasons, as well as finding out which kind of design is more likely to be repetitive and might be valuable for building up a machine learning dataset.
I interviewed 5 Autodesk Revit customers around the world. My discussion with them focused on two main topics:
- How do you organize your past documents?
- Tell me a time when you went to refer to a past project for a current project. What's the occasion? How's the experience?
1. The layout of some specific room is quite modular and have similar requirements
2. Similar construction regulations/laws(i.e. fire, noise)/ADA disability requirements
3. Personal preference/style.
1. Layout: Bathroom/ kitchen/ bedroom/ hotel rooms/ patient room and healthcare facilities/ warehouse
2. Construction details: Wall/ roof drain/ HVAC
3. Disability clearance
4. Detailed product information: i.e. type of a wall/ size of a door or tub
Intern/ Entry-level architects
Based on customers' current experience and their pain points, I created the journey map.
There are tons of opportunities for using machine learning to help to make the workflow of reusing past data more efficient.
- How might we make searching for past projects easier?
- How might we help with locating certain content?
- How might we help with extracting the information effectively?
- How might we adjust the model automatically?
MVP: Similarity Search
The goal of this project is to use machine learning to leverage architects' previous work to enhance future work and allow data to be reused in customers' iterative workflow. Considering the lack of datasets and the technical challenge, we decided to move to the direction of the similarity search as the first step. The ultimate goal is to make it easier for customers to find references from the past.
The approach is divided into three stages: Room Classifier, Layout Search, Layout Recommendation
STAGE 1: ROOM CLASSIFIER
The first phase is a tool that helps users identifying room layout. With this tool, customers could quickly check what types of rooms are on different projects. This is the starting point of exploring the capability of the current machine learning algorithm.
In this phase, we are only using users' own directories for machine learning to identify room types.
Collaborated with a machine learning engineer intern, this tool was fully developed. Here's a demo of how it works.
STAGE 2: LAYOUT SEARCH
The next phase is a tool that helps users identifying room layout and allow them to search specific room layout from previous projects. Then the users could choose the reference and open the file.
Searching and filtering function would be available. Once the user specifies what to look for, this tool will list all the similar floor plans for the user to refer to.
I designed several wireframes, run feedback sections, and did quick iteration.
One question I encountered through wireframing was that how would users decide which file to refer to. So I conducted usability testing to understand how might we help users to select from the results more effectively.
The tests were conducted through UserTesting.com. Participates were asked to finish the task by the given interactive prototype and answered a few questions afterward. The tests were separated into two sections with two prototypes.
The first prototype I used for testing had a relatively simple interface without filters. After entering the type of room, the results are shown in thumbnails with project names and dates. The users could preview a bigger thumbnail before choosing.
- Filters including project type and dimension might be helpful but not necessary, as it’s how the layout looks like that matters.
- Users would want to see more from the preview, rather than just a bigger thumbnail.
The second prototype has several filters and users have more capacity in the preview. List view is also available if the users want to see detailed information.
- Filters are helpful, but not much. They should be organized by level like a tree structure.
Based on the feedback from the usability testing. I iterated my design and made it in higher fidelity. The demo shows how would the experience from entering the room to filter and narrow down the result and then open for reference works.
STAGE 3: RECOMMENDATION TOOL
The desired state is this tool would be initiative suggesting similar designs to users during their design process. So instead of searching for reference, users could check the recommendations and bring in the design.
This animation below shows how it would work.