Architects play a pivotal role play in the sustainable development of the built environment, shaping spaces that minimize environmental impact while enhancing human well-being. However, due to the escalating complexity of design requirements, designers are becoming underequipped to solve today’s design challenges. Solving such problems involves balancing multiple different co-dependent factors, such as spatial configuration, daylight satisfaction and embodied carbon. Currently this is addressed through inter-disciplinary collaboration where different parties focus on their own specialisation. As a result, a significant gap exists between design and building performance analysis leading to underperforming and expensive solutions.

An artificial intelligence-powered platform could aid in the design process by bridging the gap between different disciplines. Such a platform could ensure that all aspects of a project are considered holistically, reducing the risk of conflicting decisions that cause lacking performance and unnecessary expenses. Simply consolidating information, as done with building information modelling (BIM), is not sufficient. Instead, the platform must provide performance-driven recommendations, offering deeper insights when conflicts between different disciplines arise.

This work aims to develop a framework through which deep learning methods can be applied to create floorplans informed by performance-based criteria and user guidance. Since functionality is heavily dependent on site-conditions and requirements laid out by the client, the relationship between form and function must be uncovered dynamically. To achieve this, inspiration will be drawn from the field of reinforcement learning which allows the training of a neural network without the use of training data.

Existing content-based image retrieval models work well for natural photos, but not for images of architectural floor plans.
Previous work on floor plan retrieval has focused on graph-based methods, rather than image-based floor plans.
Training a CNN-based representation learning framework on segmented floor plan images with standard image augmentations does not result in semantically meaningful retrievals.
This work shows that a CNN-based representation learning model can learn features for retrieving floor plans that have similar graphs given the right training signal. Two methods were investigated here: GeomPerturb, a data augmentation that perturbs the underlying geometry of a floor plan, and a weakly supervised method with labels based on the graph edit distance between a pair of floor plans. The results show that while GeomPerturb learns representations that are correlated with the floor plan graph, training with GED labels leads to better retrievals both in terms of the floor plan graph and with respect to room shapes.