A new preprint on arXiv tackles a persistent bottleneck in building design: the high computational cost of computational fluid dynamics (CFD). The authors propose that deep learning surrogates could provide faster predictions, but note that generalization across varying room dimensions remains a key challenge.

The study compares two families of geometry-conditioned neural networks: canvas-based and coordinate-based. These approaches differ in how they encode the geometry of indoor spaces, which may affect their ability to handle rooms of different sizes and shapes.

The abstract does not yet disclose results or a clear winner, so the practical advantages of either architecture remain to be seen. The work is positioned as a step toward real-time prediction tools for iterative design workflows.