NVIDIA Earth2Studio Enables Custom Batched Weather Forecasting

NVIDIA Earth2Studio has been utilized in a tutorial to construct custom, batched ensemble weather forecasting workflows. This process involves installing the necessary Earth2Studio components, ensuring compatibility with an existing CUDA-enabled PyTorch environment within Google Colaboratory. The tutorial guides users through loading the FCN (Fully Convolutional Network) prognostic model, a type of deep learning architecture commonly used for image segmentation and, in this context, for weather prediction tasks. It also details how to retrieve atmospheric initial conditions from the Global Forecast System (GFS), a widely used numerical weather prediction model developed by the National Oceanic and Atmospheric Administration (NOAA).

The workflow includes the implementation of a custom wind-power diagnostic. This diagnostic tool converts 10-meter wind components, a standard meteorological measurement, into turbine capacity factors. Turbine capacity factor is a metric that represents the ratio of actual energy generated by a wind turbine to its maximum possible output over a given period. Furthermore, a variable-scaled perturbation system is introduced. This system applies physically appropriate noise amplitudes to different atmospheric variables, such as temperature, pressure, and wind speed, while crucially maintaining an unperturbed control member for comparison. This perturbation is essential for ensemble forecasting, as it generates multiple forecast scenarios to capture uncertainty.

Leveraging Earth2Studio's specialized APIs, including its low-level iterator, coordinate-mapping, batching, and Zarr APIs, a custom ensemble execution pipeline is constructed. The Zarr API is a format for storing chunked, compressed N-dimensional arrays, facilitating efficient data handling for large scientific datasets. The pipeline is designed to write forecast and diagnostic fields to a coordinate-aware data store, ensuring that spatial and temporal information is preserved. The accuracy of these forecasts is then verified against GFS analyses. Verification metrics employed include latitude-weighted Root Mean Square Error (RMSE), which measures the average magnitude of the errors, considering the Earth's curvature; fair Critical Success Index (CRPS), a probabilistic score that measures the accuracy of a probabilistic forecast; ensemble spread, which quantifies the dispersion of the ensemble members; and spread-skill ratios, which assess the reliability of the ensemble by comparing its spread to its error.

Finally, the tutorial illustrates methods for visualizing the ensemble uncertainty. This includes generating spatial maps to show the geographical distribution of forecast variables, geopotential-height spaghetti contours to depict variations in atmospheric pressure levels, and point-based fan charts for time-series forecasts. The visualization also covers wind-capacity-factor forecasts and lead-time skill curves, which demonstrate how forecast accuracy changes with the forecast horizon. This comprehensive approach allows for a detailed understanding of potential weather outcomes and their associated uncertainties, enabling more informed decision-making in sectors reliant on weather predictions, such as renewable energy and disaster preparedness.