Constraining AI-based Parameterizations for GCMs using Long-term Observations

The representation of processes occurring on scales finer than typical resolutions used in weather and climate models constitutes the most significant source of uncertainty in climate projections. Such processes include cloud microphysics, turbulence, boundary layer dynamics, and tracer transport. Several AI-based parameterizations have been developed to address the representation of subgrid scale physics. However, AI models trained only on simulated data may lack the stability and robustness required for reliable climate predictions and may not adhere to physical laws. Unless constrained by a set of physically plausible outcomes, these models risk becoming fragmented with different datasets and architectures, leading to a proliferation of parameterizations that do not effectively reduce uncertainty. Here we present generative methods to combine high resolution simulated data, climate reanalyses, and long-term observations to generate constrained subgrid scale parameterizations. Specifically, we focus on the representation of vertical wind velocity, with implications for cloud microphysics and turbulence. Our results indicate that AI models trained only on simulated data do not fully capture the subgrid variability in vertical wind, leading to biases in cloud representation. We will present new results characterizing the accuracy of constrained AI models with respect to observational datasets using novel statistical techniques.