obsweatherscale: a Python library for ML-based probabilistic interpolation and downscaling of surface weather fields

This poster presents the library obsweatherscale and showcases it. 

obsweatherscale is an open-source Python library for machine learning-based probabilistic interpolation and regression of surface weather variables using Gaussian Processes (GPs). Built on GPyTorch, the library provides a modular and extensible framework for constructing neural-augmented Gaussian Processes models that incorporate trainable mean and kernel functions that accept arbitrary input features.  Key features include plug-and-play data transformations, support for uncertainty quantification,  GPU acceleration via the deep-learning framework PyTorch (https://pytorch.org/projects/pytorch/), as well as training and inference routines.

The library was developed to generate high-resolution wind maps over Switzerland, where complex alpine terrain and sparse observations challenge traditional methods. The aim was to create detailed wind maps that closely align with observations. Leveraging comprehensive orography maps, topographic descriptors, and numerical model outputs, we use obsweatherscale to downscale hourly ICON reanalysis wind fields from 1 km to sub-kilometer resolution integrating the wide variety of predictors. This showcases the model's ability to improve spatial detail and observational consistency, as well as provide calibrated uncertainty estimates.

It may be useful to note, that obsweatherscale can be applied to any surface parameter and additional applications: Beyond wind downscaling, obsweatherscale generalizes to a range of meteorological applications, including bias correction of model outputs and probabilistic spatial interpolation of observational datasets.

 

References: 

Lloréns Jover, I., & Zanetta, F (2024). obsweatherscale: observation-conditioned ML downscaling of surface weather fields. GitHub repository: https://github.com/MeteoSwiss/obsweatherscale 

Zanetta, F., Nerini, D., Buzzi, M., & Moss, H. (2025). Efficient modeling of sub-kilometer surface wind with Gaussian processes and neural networks. Artificial Intelligence for the Earth Systems.