Spatial graph learning for fresh air duct topology and size prediction in general hospitals

Designing fresh air duct systems for general hospitals is challenging because ventilation requirements must be coordinated with complex clinical functions and highly variable spatial layouts. In conventional practice, duct routing and sizing depend largely on the experience of HVAC engineers and are usually developed after the architectural layout has taken shape. Consequently, architects receive limited quantitative information about duct routing and spatial requirements while room arrangements are still being adjusted. This is particularly relevant in hospital buildings, where duct distribution must be coordinated with infection-control requirements, ceiling space, indoor environmental quality, and maintenance access. This study develops a two-stage graph-based learning framework to provide early predictions of fresh air duct topology and duct segment sizes from architectural spatial information. In the first stage, each functional unit is represented as a spatial graph, and a graph neural network (GNN) predicts duct connections between candidate pairs of adjacent spaces using spatial adjacency, connectivity, and node attributes. In the second stage, an Extra Trees classifier predicts the standard size of each duct segment from the directed duct network, the attributes of the spaces served by terminal ducts, and downstream demand proxies. This stage is tested using the duct segment topology extracted from the BIM models, rather than the space-level connections predicted in Stage 1. The experimental results show that spatial layout graphs built on spatial adjacency, connectivity, and attributes can support the prediction of the topology and duct segment sizes of fresh air duct networks at the early design stage in general hospital buildings. The model that predicts fresh air duct connections between spaces achieved a Precision of 0.633, Recall of 0.745, F1-score of 0.684, and AUC of 0.900. For the prediction of duct segment sizes, the Top-1 accuracy for standard duct sizes reached 0.414, and the MAPE of the cross sectional area was 0.339. When the top three candidate standard sizes were output, the coverage rate of the true standard size increased to 0.593, and the MAPE of the closest candidate cross sectional area decreased to 0.114. These results indicate that spatial layout graphs can provide an effective basis for predicting fresh air duct network topology and candidate duct sizes at the early design stage in general hospital buildings. Future work should focus on improving the identification of terminal fresh air demand from spatial attributes, which is expected to further enhance the accuracy of duct size prediction for the entire network.