Generative Models and the Challenge of Representing Vernacular Architecture in Hot-Arid Climates

Generative layout models are proposed as general instruments of architectural design, yet every large public floor-plan corpus on which they are trained originates in a narrow band of temperate and continental climates and in an extroverted, room-hub dwelling typology. This paper asks what follows when such a model is applied to hot-arid contexts whose historical typology is the introverted courtyard house. The contribution is addressed to the learning system rather than to the building: what a plan representation is able to encode, what a conditioning mechanism is able to constrain, and what a corpus-referenced metric is able to reward. The courtyard house is the test case, chosen because it makes those three limits measurable. A documentary audit of four principal corpora, together exceeding two hundred thousand annotated plans, establishes that none documents an origin in a hot-arid or semi-arid Köppen zone and that no unroofed interior court appears as a distinct spatial class in any of their published label vocabularies; two corpora released since, one of them explicitly assembled for geographical diversity, are shown not to alter the finding. We then argue that the resulting limitation is not a data shortage but a failure of encoding. Of five requirements of the typology, only one is a shortage of examples. The courtyard is a void that organises the plan yet carries no enclosure attribute; the constraint that habitable rooms do not open to the street cannot be written where the exterior is not a node; the privacy gradient is a property of topological depth that adjacency conditioning does not express; and seasonal migration between winter and summer quarters is absent from a static plan representation altogether. A model can therefore be enlarged with local data and remain structurally blind to what makes the typology work. We introduce the Corpus-Context Distance protocol, a five-dimension diagnostic applied before deployment that yields a disclosure statement for the design record, and we specify what a corpus adequate to hot-arid architecture would have to encode. The study is documentary and conceptual; its propositions are stated so that they can be refuted.

کلیدواژه ها:

generative design ، floor plan generation ، training corpora ، courtyard house ، hot-arid climate ، spatial privacy ، algorithmic bias