Synthetic data generation for machine learning

The first step; Setting the framework for synthetic data generation of indoor spaces

The initial steps of the collab aimed at investigating how Parametric’s generative capabilities could be used to create large data sets of indoor geometries, tailored to the purpose of training machine learning algorithms developed by Treble. As a pilot project, a set of 1000 meeting rooms was generated.

With a diversity of geometric configurations, such as the shape of the room, placement of furniture and acoustic elements the generative design algorithms developed on Parametric’s side proved successful at creating a consistent dataset of watertight 3D-geometries – the latter being a strict requirement for Treble’s algorithms to work optimally.

Having been a success, the pilot helped setting the framework for the continued collaboration; Not least in learnings on how to approach geometry generation further down the line – when all aspects of geometry and metadata exponentially grew in complexity.

‍

Parametric visiting Treble's Office in Reykjavik, September 2023. From the left: Steinar, Jesper, Erik (Parametric), Simon (Parametric), Finnur and Lena.

‍

The second step; Expanding to more complex typologies

Following the meeting rooms, the algorithms were pushed and in part, redefined, to generate more complex types of spaces.

Residential apartments were decided on as the next logical step, due to its many unexplored applications in terms of acoustics. This entailed developing an algorithm understanding different functions and room types typically found in dwellings – as well as the interconnection between them.

A divide-and-conquer approach was used to sub-divide seemingly unmanageable problems into smaller tasks that could be solved individually, namely:

‍

1. Creating a set of area-defined boundary curves with ancillary facades and entrance points that resemble a real-world dataset of apartments.
2. Dividing an arbitrary boundary curve into separate spaces.
3. Furnishing an arbitrary space according to the type of room, i.e. what objects are allowed inside it and how they can be arranged.
4. Assigning appropriate sample points to an arbitrary space to be used for acoustic simulations.

‍

Example of a watertight 3D-geometry, in this case a one-bedroom apartment

‍

The third step; Initializing Parametric SDK

Moving on, Parametric continuous aiding Treble in training their and their customers’ acoustic machine learning models to produce even better results for even more environments. The collaboration looks into creating a wider catalogue of data sets such as restaurants, educational spaces, office spaces, theatres etc. It seems to us at Parametric that almost all architecture could be recreated generatively if only you know what guiding principles to look for.

As a direct result of the ongoing collaboration with Treble, we at Parametric are now looking into developing our own SDK, enabling calling vast amounts of 3D-geometry and data as efficiently as possible directly through code.

We are incredibly excited for the continued partnership with Treble and eager to unveil future progress. Thank you Treble for a fruitful endeavour so far and for sharing your enormous expertise!

Sign up for our newsletter to stay posted about how our collaboration with Treble advances and to learn more about the SDK development.

Read more about Treble Technologies and their acoustic suit here.