AEC Tech 2018 – LunchBoxML Workshop Examples

David Stasiuk and I instructed a 4-hour workshop on Machine Learning as part of AEC Tech 2018 – an annual conference hosted by Thornton Tomasetti. The workshop focused on demonstrating architectural use cases for machine learning in the context of our open source LunchBoxML tools. Our course examples are now available on the LunchBoxML Bitbucket repository.


One set of examples focused on the use of regression tools to derive prediction curves based on an input set of data points. This workflow using nonlinear regression was then introduced to determine a curve representing the face of a point scan slice.

Nonlinear regression tools were applied to determine a curved edge within a point scan slice.

Gaussian Mixture

Another example involved the use of a Gaussian Mixture algorithm to sort panels based on attributes including size, flatness, and edge lengths. This workflow allows for the rapid clustering of similar panels in order to facilitate fabrication workflows.

Workshop participants apply a Gaussian Mixture algorithm to sort and cluster panels on a complex surface.
Another application of the Gaussian Mixture is in its use to cluster other complex data sets, such as point scans, which can aid in feature identification.

Naive Bayes

The workshop concluded with an example for using a Naive Bayes classifier to help a user predict facade “type” assignments. This example uses a sample data set stored in Excel indicating facade attributes (orientation, size, program) and specifies the use of a facade type (abstractly defined as “low”, “medium”, and “high”). Within Grasshopper, a user is then able to define a massing volume to test the classifier’s ability to determine the likelihood of a facade type to be selected.

A Naive Bayes classifier used to determine facade types on 3D massing using sample input data.

Workshop Conclusions

The AEC Tech 2018 workshop was the first public workshop we have given focused on uses cases for LunchBoxML.┬áComputing topics, like machine learning, are experiencing a great deal of hype within the construction industry. In many cases we have observed that the technology is positioned as a solution in search of a problem (sometimes called ‘solutioneering’). To avoid falling into this trap, our workshop covered a variety of ideas for how to position these tools within real-world design workflows. Our goal was to discuss the conceptual basis for the algorithms included in LunchBoxML while demonstrating their use through a set of simple, practical examples.

We hope that this gave participants a foundation to build from as they seek to apply ML in their design work.