Emulation of sub-grid physics using stochastic, vertically recurrent neural networks

Machine learning (ML) has the potential to reduce systematic uncertainties in Earth System Models by replacing or complementing existing physics-based parameterizations of sub-grid processes. However, after decades of research, ensuring generalization and stability of ML-based parameterizations remains a major challenge.  We aim to minimize both epistemic and aleatoric sources of uncertainty via physically inspired, vertically recurrent neural networks (RNN) which offer key benefits such as parametric sparsity and efficient modeling of non-locality in a column. To address aleatoric uncertainty, we furthermore incorporate stochasticity and convective memory into the ML architecture. We present preliminary results using the ClimSim framework, where the physically inspired ML framework replaces a superparameterization in a low-resolution climate model.