Interpretable latent modelling of canopy geometry for aerodynamic drag inference

We develop a data-driven approach to infer aerodynamic drag directly from canopy geometry by learning interpretable low-dimensional latent representations of canopy configurations, without resolving the flow field. A PixelCNN-based variational autoencoder with latent disentanglement regularisation and auxiliary observable regression is used to identify latent factors that encode the geometric features most relevant to drag. Latent traversals and mutual information analysis are employed to quantify the physical relevance of individual latent dimensions and to assess how modelling choices affect information retention. Applied to laboratory measurements of heterogeneous canopy arrays, the learned latent space organises canopy configurations according to their aerodynamic impact and enables accurate drag prediction. The results demonstrate that physically informed latent modelling provides a compact and interpretable pathway for linking complex canopy geometry to aerodynamic drag.