Editor's Highlights are summaries of recent papers by AGU journal editors.
sauce: Journal of Advances in Earth System Modeling
Modern Numerical Weather Prediction (NWP) models can predict weather up to 15 days ahead with high accuracy. These forecasts save lives by preparing for extreme weather events, and are of great economic value, being widely used in the energy, transportation, agriculture, and leisure industries.
It took 50 years to build and incrementally refine the sophisticated computer models needed to make such accurate predictions. In just a few years, machine learning (ML) approaches have matched and in some cases surpassed the skill of these traditional models.
In the face of this paradigm shift, we recognize the need to compare ML models with each other and with traditional models in a fair and reproducible way. Rasp et al. [2024] They address this need. Drawing on 50 years of experience in the NWP community, they present a framework for evaluating both ML and NWP forecasts. Called WeatherBench 2, the framework outlines standard metrics for comparing forecasts and provides open-source code and data to facilitate the process.
The community also recognizes that there is room for improvement in both ML and NWP forecasting, including longer interseasonal and subseasonal forecasts, probabilistic approaches, and targeted tuning for specific use cases. WeatherBench 2 lowers the barrier to entry by clearly defining the forecasting problem and providing easily accessible training data for ML models. In this way, talented minds from around the world can jump start research in these and many other areas, promising better weather forecasting for all.
Citation: Rasp, S., Hoyer, S., Merose, A., Langmore, I., Battaglia, P., Russell, T., et al. (2024). WeatherBench 2: Next-generation data-driven global weather model benchmarking. Journal of Advances in Earth System Modeling16, e2023MS004019. https://doi.org/10.1029/2023MS004019
—Hannah Christensen, Associate Editor James
