weathernext

Bindings for WeatherNext ensemble forecast models.

WeatherNext models (https://github.com/google-deepmind/weathernext) are JAX functions that map initial conditions to an xarray forecast and sample one ensemble member per random key. Wrapping such a function in WeatherNextPredictor makes it a stochastic probly predictor: every call draws a fresh member, and representer(predictor, num_samples=...) collects an ensemble and returns one NumpySample per forecast variable, ready for quantification.

Note that only sampling-based consumption is supported: probly’s model transformations (e.g. dropout) cannot be applied because the wrapped model is an opaque function without a torch or flax module tree, and training-time methods (e.g. swag) are unavailable since only final weights are released.

Classes

WeatherNextPredictor

Wraps a WeatherNext forecast function as a stochastic probly predictor.

WeatherNextRepresenter

Representer collecting a WeatherNext ensemble into per-variable array samples.