Title: Towards Physically-Consistent, Data-Driven, and Interpretable Models of Convection
Abstract: Data-driven algorithms, in particular neural networks, can emulate the effect of sub-grid scale processes in coarse-resolution climate models if trained on high-resolution climate simulations. However, they lack interpretability, may violate key physical constraints, and make large errors when evaluated outside of their training set. In this seminar, I will share recent progress towards overcoming these three challenges. First, machine learning interpretability tools can be tailored to stabilize data-driven models of convection and confirm that they behave consistently with observations. Second, physical constraints can be enforced in neural networks, either approximately by adapting the loss function or to within machine precision by adapting the architecture. Third, as these physical constraints are insufficient to guarantee generalizability, I additionally propose to physically rescale the inputs and outputs of machine learning algorithms to help them generalize to unseen climates. Overall, these results suggest that explicitly incorporating physical knowledge into data-driven models of climate processes may improve their performance, stability, and consistency.