Earth Systems Models (ESMs) are the primary tool used by scientists to understand earth system dynamics and variability and make future climate projections. The accuracy of ESMs has increased significantly in recent years due to advances in numerical modeling and computational tools, allowing them to be run with more precision at increasingly higher resolutions on high performance computers (HPC). Earth system processes are primarily described by non-linear differential equations that interact with other complex and non-linear components on differing spatial and temporal, resolved and unresolved scales. The increase in the degrees of freedom with resolution implies that solving these multi-scale equations for high resolution models becomes computationally prohibitive, even on leadership-class HPC. Moreover, the unresolved processes also account for a large source of uncertainty in ESMs, as they are approximated from the resolved scales through parameterization schemes.
Recently, researchers have been exploring the use of AI to aid parametrizations of a wide range of sub-grid scale processes including but not limited to radiation, convective processes, aerosol and cloud microphysics, atmospheric chemistry, ocean physics, and land processes. ML can be used to create new parametrizations, or improve current ones, by training a statistical model to effectively emulate a complex physical phenomenon. ML techniques are capable of learning and accurately simulating nonlinear functions when trained on large datasets. Once trained, these models can provide fast calculation of sub-grid scale processes, particularly on hybrid architectures. Current ML tools used for parametrization tasks include neural networks (NN), deep learning, random forests (RF), and general adversarial networks (GAN). Several studies have also shown the usefulness of hybrid approaches called physics informed ML, also known as knowledge guided ML, for incorporating domain knowledge and physical principles into the ML techniques mentioned above. The main challenges of utilizing these ML approaches include the limited explainability and uncertainty quantification of the resulting models. Alleviating these limitations would boost our confidence in the applicability of the ML models to future projections.
We aim to explore different ways ML techniques can assist in defining and representing parametrization processes across earth system modeling. Submissions related to the use of ML techniques for ESM parametrization including emulation, replacement and acceleration of current parameterization schemes are sought to further our understanding of advanced ways to address model run-time expense, accuracy, sensitivity and uncertainty.

