InsAIderKnowledge

The project InsAIderKnowledge addresses the challenging problem of conditional building generation. In this research project, 2D floor plans of Bavarian single-family houses remain the central focus and are planned as deliverables, enabling sub-stream workflows for professional planning in the AEC industry. The generated houses should be consistent with the exterior photos taken from the street or aerial photos from Google Earth. This task could be formalized as a generation problem from a single image or multi-view images. Since exterior observations only reveal part of the building geometry and provide little information about the interior organization, many different floor plans may comply with the same exterior observations. This task is therefore under-constrained. The hypothesis is that this under-constrained generation task can be guided by architectural knowledge. Contemporary AI methods will be applied to justify this hypothesis.

Relevant architectural knowledge can be sourced from several categories, including spatial design principles, architectural style, construction period, building regulations, and more implicit forms of design intentions.

Among these knowledge sources, spatial and functional principles are particularly well represented in 2D floor plans. Floor plans form the foundation of professional architectural planning workflows and encode substantial formalizable knowledge, including room adjacency, accessibility, proportions, and typical room sizes. (e.g., corridors tend to be elongated and relatively narrow, whereas living rooms tend to be square-shaped and larger.”)

Style or construction period also strongly influences the interior layout of a single-family house. e.g., in the old times, kitchens tended to be closed off, and dining rooms were isolated. But this has changed in modern design, where kitchens tend to be open to the dining and living room. Consequently, even houses with similar exteriors may result in different layouts depending on their construction period or style. Style, therefore, acts as an additional condition that helps resolve the ambiguity inherent in image-based floor-plan generation.

Building regulations are another important condition. This can vary across neighborhoods or within the same neighborhood at different times. Unlike stylistic or functional preferences, these are often hard constraints and must be satisfied for the resulting design to be considered compliant. Such hard constraints should therefore be handled differently from preference-based conditions.

Because the aforementioned conditions originate from different sources and operate at different levels of authority and abstraction, they may be mutually incompatible. For example, in the real world, contradictions often arise between human-designed regulations, especially when state-level regulations conflict with local regulations. Conditions should first be distinguished as hard constraints or soft preferences. In addition, rules within the same category may have different levels of authority. A conflict-resolution mechanism is therefore required to determine which condition takes precedence when constraints are incompatible.

While some architectural knowledge can be represented explicitly as rules or constraints, other forms of design knowledge are fuzzy, subjective, or only partially formalizable, such as notions of “awkwardness,” “spaciousness,” or the “quiet side of the building.” Such intentions can sometimes be approximated through explicit formalization. For example, room types such as bedrooms may be categorized as “quiet side” in a predefined taxonomy, and graph-based relations can then be used to represent this knowledge. However, such handcrafted representations work only at a limited scale and capture only a small portion of the underlying design intention. This motivates a hybrid approach in which explicit constraints are used to capture well-defined requirements, while learned representations are employed to capture softer, less easily formalized design intentions.

A separate methodological question concerns the representation in which the generation should take place. 2D floor plans can be easily lifted to 3D buildings with post-processing, such as by injecting default vertical parameters (e.g., window height). In addition to directly starting from 2D floor plans, methods that first directly generate 3D and are post-processed to 2D will also be studied.