Creating custom weather forecasting models with nvidia's earth2studio

A tutorial has demonstrated how to build a custom batched ensemble weather forecasting system using NVIDIA's Earth2Studio. The process involves installing Earth2Studio while preserving Colab's existing CUDA-enabled PyTorch environment, loading a FCN prognostic model, and retrieving atmospheric data from GFS. A custom wind-power diagnostic was implemented to convert 10-meter wind components into turbine capacity factors, alongside a variable-scaled perturbation system that applies physically appropriate noise to different atmospheric variables. The tutorial also highlights the use of Earth2Studio's low-level iterator, coordinate-mapping, batching, and Zarr APIs to construct an ensemble execution pipeline. The workflow includes retrieving GFS analyses for verification, calculating metrics like latitude-weighted RMSE and fair CRPS, and storing results in a structured dictionary. Visualization tools such as spatial maps, spaghetti contours, and fan charts were used to illustrate ensemble uncertainty. The final Zarr dataset, accessible via Xarray, preserves the complete ensemble structure and allows for further analysis. The tutorial emphasizes the flexibility of Earth2Studio, enabling users to extend the workflow by modifying models, data sources, or diagnostics without redesigning the entire pipeline.