EESS talk on "Improving AI-Driven Subgrid Parameterizations in Climate Models Using Long-Term Observational Data’"

Small-scale processes like cloud formation, turbulence, and boundary layer dynamics are difficult to represent in global climate models and constitute a major source of uncertainty in climate predictions. Parameterizations based on artificial intelligence (AI) offer a promising solution for representing these subgrid processes. These AI models are typically trained on high-resolution simulations, but relying only on simulated data can lead to stability and consistency issues. In this seminar, I will present a new approach that combines high-resolution simulations, climate reanalysis, and long-term observational data to create physically-constrained AI models for subgrid processes. Specifically, we focus on vertical wind velocity (W), a crucial factor for cloud formation and turbulence. Our research shows that models trained only on simulated data struggle to accurately represent subgrid variability in W, leading to biases in cloud formation rates. By incorporating observational constraints into AI models and applying them to climate reanalysis data from 1980 to 2023, we reveal significant trends in W variability, particularly in oceanic regions. These trends, showing an annual increase of up to 1%, suggest a more turbulent atmosphere with increased cloud formation, which could have important consequences for future climate predictions.