Wind fields: a ‘topological’ adjustment helps a neural network to better reproduce areas of strong wind

By incorporating penalties for local maxima and gradients into the training process, researchers have improved the super-resolution of 150 out of 168 wind fields, with topological deviations reduced by between 25 and 31 per cent.

The super-resolution consists of reconstructing a detailed version of a low resolution meteorological field. Arjun Dadhwal and Brian Summa have taken over a convolutive network already trained for this task on wind fields, and have refined it with loss functions inspired by topology: they concern the speed, the intensity of the gradients, the areas of strong wind and especially the local maxima, those peaks that matter for wind and for the anticipation of violent phenomena.

Preserving structure rather than average

The method is based on the Topology ToolKit and forces predicted values at key points in the structure of the real field. In a 168-field bench (160 × 160-point sub-region), topological distances at the reference field decreased by 25 to 31% from the baseline model, and 150 fields out of 168 were better reconstructed. The PSNR gain, 1.30 dB, is only half that of a conventional refining: the authors deliberately favour structural fidelity over average values. This is the extended version of an accepted poster at TopoInVis Connect 2026, published on arXiv on October 15.