Building design: a neural network predicts a room’s air quality and temperature in real time, regardless of its size

To replace CFD simulations that are too slow during the sketching phase, two researchers are comparing two ways of teaching a U-Net the geometry of a part.

Verifying that an air conditioning supply will bring fresh air and a correct temperature to an entire room usually involves computational fluid dynamics (CFD), which is too computationally expensive to keep up with the rapid iterations of architectural design. Deep learning substitutes go much faster, but struggle to remain reliable when the dimensions of the room change.

Shaofan Wang and Philipp Geyer compare two geometry-conditioned models, both based on U-Net autoencoders that predict the airflow field and the temperature field separately. The first, called "canvas" (CaNN), explicitly inscribes the size of the room by delineating the occupied area in a fixed-format canvas; the second (CoNN) passes through channels of standardized coordinates.

Tested on simplified parts of varying heights and lengths, the two models are not equal: the canvas approach is more in line with the results of the CFD. The authors see it as a tool to evaluate comfort and ventilation in real time from the sketch, and to orient the design towards performance.