Building a Hierarchy of Hybrid, Neural Network Parameterizations of Convection

Computational resources currently limit climate models to spatial resolutions of O(100km) when run for time periods relevant to societal decisions, e.g. 100 years. Therefore, climate models rely on semi-empirical models of subgrid-scale cloud processes. When designed by hand, such convective parametrizations are unable to capture the complexity of convective processes and cause well-known biases, such as distorted extreme precipitation events and low-frequency variability in the Tropics. Alternatively, machine-learning algorithms trained on high-resolution global climate simulations can faithfully emulate convective processes and reduce convective biases by replacing sub-grid scale parametrizations used in climate models. In particular, neural-network parametrizations of convection emulate convective processes with high accuracy and are fast to evaluate once trained. However, neural network parametrizations are in their infancy: their performance and physical consistency must be improved if they are to be used operationally for climate modeling. In this talk, we ask: How can we best combine the complementary advantages of physical models and neural network-approaches?

We design a hierarchy of four neural networks to emulate convective processes in aquaplanet and real-geography simulations using the Super-Parameterized Community Atmosphere Model: (1) A phenomenological model of convection with fixed parameters where we train the neural network to de-bias the model, (2) the same phenomenological model with free parameters that we target with the neural network, (3) a neural network whose architecture is modified to conserve physical invariants to satisfactory numerical precision, and (4) an unconstrained neural network. The performance of each model is measured via mean-squared error of the convective temporal tendencies, while their physical consistency is measured via (1) how much the conservation equations are violated and (2) the model’s performance in out-of-training conditions (i.e., its ability to generalize).