Incorporating physical knowledge in machine learning parameterizations of convection

Abstract:

There has been recent demonstrations of the potential of machine learning approaches to convective parameterizations, informed by high-resolution cloud resolving models, using unsupervised learning. However an inherent assumption in machine learning applications is that the underlying process distribution does not change; an assumption intenable with climate change. In addition, climate model oriented parameterizations need to exactly conserve mass and energy conservations - which has proven untrue using neural network approaches.
We will here elaborate on methodologies to tackle those two issues: 1. generalization outside of the initial distribution and 2. Exact conservations. The new physics-abiding models demonstrate more skill and require less data for training this opening the door for their use in climate projections.

There has been recent demonstrations of the potential of machine learning approaches to convective parameterizations, informed by high-resolution cloud resolving models, using unsupervised learning. However an inherent assumption in machine learning applications is that the underlying process distribution does not change; an assumption intenable with climate change. In addition, climate model oriented parameterizations need to exactly conserve mass and energy conservations - which has proven untrue using neural network approaches.
We will here elaborate on methodologies to tackle those two issues: 1. generalization outside of the initial distribution and 2. Exact conservations. The new physics- abusing models demonstrate more skill and require less data for training this opening the door for their use in climate projections.