Description:

Over the past five years, the field of data-driven weather forecasting – leveraging large, deep learning models trained on a massive corpus of historical weather data, typically a multi-decadal reanalysis such as ERA-5 – has rapidly developed and matured, with both private companies and public agencies running these tools operationally as of late 2023. However, despite this rapid pace of progress, there is still a substantial body of active research and development aimed not only at improving these tools, but also working to evaluate and validate them for applications beyond simple medium-range, global-scale weather forecasting.

The goal of this session is to highlight the developments and achievements within the field of data-driven weather modeling, with a keen focus on emerging challenges, performance evaluation, questions, capabilities, and opportunities:

• Where are the barriers towards developing – and operationalizing – data-driven weather models?
• What model architectures and problem formulations are emerging as candidates for further advancing the paradigm of data-driven weather modeling?
• How will forecast products derived from data-driven weather models co-exist with the extant and continually evolving catalog of operational NWP systems? Do we anticipate a scaling issue?
•  To date, data-driven weather models have excelled at forecast applications ranging from precipitation nowcasting, global medium-range forecasting, and S2S forecasting. Where are there untackled forecasting problems where this technology may prove even more efficacious?
• How are these models performing with precipitation events in short-range, medium-range, and long-range forecasts? 
• Are traditional skill metrics still adequate for evaluating “non-physical” data-driven weather models? If not, how should these tools instead be evaluated?
• Many state-of-the-art data-driven weather models are trained on ERA-5 and evaluated using the WeatherBench2 framework. Are there extensions - such as new training datasets, new baseline/evaluation problems, etc – which the community should consider when continuing to develop these tools?

We especially welcome perspectives on how the broader AI and NWP communities might come together to further accelerate this weather forecasting revolution. Prospective authors may consider submitting abstracts that provide a performance assessment of predictions from their data-driven ML models which include comparisons to benchmark predictions from single deterministic NWP models, and/or an ensemble of NWP models. This session will encourage research into the use of purely data-driven ML models to generate predictions with skill comparable to that from state-of-the-art NWP models used by operational meteorologists worldwide.

Please specify which of the sub-session titles you wish to contribute to:

• Making it rain — precipitation forecasting with AI/data-driven weather models
• Beyond RMSE: Dynamical and physically-based evaluations of AI/data-driven weather models
• Leveraging data-driven weather models for very large ensembles
• Diffusion modeling applications for weather forecasting