Analytics, machines, and our summer so far…

We are in the heat of the summer season and we have invited summer interns to work with us to experiment with new concepts and technologies. In 2016, we focused on rapid development of focused problem-first workflows and published our content as an open source project.

This year, Ally and Nazanin have joined us for a summer internship and we are beginning to explore at some challenging concepts with regards to the uses of analytics and machine learning in design…

Interior Design Analytics

Data visualization and analysis has been an important part of our service capabilities for some time now. We often use a data-centric approach to analyze building models to gain insight into quality and also identify trends in certain attributes such as the building program. We have developed a custom relational database for capturing Revit model data for analysis purposes and are actively deploying it with our clients.

Internally, we decided to put this new database to the test by harvesting data from interior design models. This has allowed us to identify improvements to the process and also understand the kind of data we are collecting. Ally solicited Revit models from her recent design studio at Virginia Tech. Given that the entire class had produced designs based on the same program for a co-working space, we thought it would be interesting to look for variations in how different designers tackled the same problem.

AreaStudy

SpaceStudy

Ally used the database to prototype SQL queries to compare interior design features such room attributes, lighting content, and furniture types. PowerBI was used to visualize this data and produce interactive dashboards. This data also became a valuable resource to inform other data-driven studies….

Teaching the Machine

If you have been to a recent conference you would have been hard pressed to make it through the event without hearing the term “Machine Learning”. This new dark art has captivated the imaginations of many technologists in recent years. In short, machine learning is a computer science concept for ‘teaching’ a computer how to solve a problem without explicitly writing a program to solve it. This concept is useful when we need the computer to help us make predictions or determine patterns in complex data. Facial recognition and predicting customer purchase behavior are some of the more common machine learning applications that you probably interact with every day on social media and online stores.

In architecture, Daniel Davis’ recent research at WeWork proposes one of the more tangible use cases for how this idea can inform building design and operations. They implemented an approach to collect data about conference room usage and used machine learning to make predictions so they could better calibrate their future work environments.

We decided it would make for an interesting exploration to start to build our own machine learning toolkit within parametric modeling software and create some simple use cases.

In one scenario, we wanted to teach the computer to make a prediction of what furniture should be specified to finish a space using prior Revit models as the learning data. Nazanin used space data that Ally had mined from the interior design models and implemented a Naive Bayes classifier to predict the furniture needs of new space.

ML_FurnitureContent
Based on space conditions, the algorithm predicts furniture content needs based on a history of past Revit models with similar programs.

A similar study was conducted focused on a assigning facade types with different performance characteristics based on previous building designs. We created a sample data set that considered characteristics such as orientation, program, and facade area in relation to a conceptual understanding of facade types with different performance characteristcs. We then fed in several building masses into a Grasshopper definition analysed different faces to predict the facade type.

ML_Facades
Based on data about facades on different orientations and building programs, the computer will predict the facade types of a new block massing.

What is interesting is that the process does not use any form of environmental or energy simulation to arrive at a result: Instead the process looks at a history of design conditions to predict what should be assigned to the facade. Of course, if a data set demonstrates a history of poor design, the computer will have learned the wrong behavior… even with learning machines, it is still “garbage in, garbage out.”

What’s Next…

We’re still thinking through the possibilities of these early ideas. Like with last year’s studies, we hope to provide a more thorough report by the end of the summer.

Stay tuned….!