Verification, diagnostic, and interpretability of AI models for weather forecasting

Convener: Zied Ben Bouallegue | Co-conveners: Jochen Broecker, Philine BommerECSECS, Romain PicECSECS

Weather forecasting based on AI models is now part of our operational and research landscape. AI-based weather forecasts show improved skill with verification measures such as the root mean squared error when compared with NWP models. However, proper and in-depth assessment of strengths, weaknesses, and properties of these models is still ongoing. This session aims to gather contributions advancing the assessment of AI-based weather forecasts.

This session welcomes contributions on the following topics with applications to AI weather models:

• Benchmarking activities (e.g. datasets, intercomparison projects, comparison with NWP forecasts)
• Verification methodology (e.g. spatial verification methods, scoring rules, or innovative approaches)
• Diagnostics of forecast realism and potential forecast artifacts
• Forecasting extreme events, predictability, and other properties (e.g. fairness)
• Interpretability of AI weather models, e.g. XAI methods.

Contributions covering theoretical, methodological, applied, or operational aspects are equally welcome.