CAM6 climate model: a neural network replaces the deep convection scheme and corrects the tropical rainfall cycle

Trained on a global model that simulates thunderstorms, the parameterisation has been running for over ten years without drifting and reduces biases in both light and extreme rainfall.

Deep convection, which forms large storm clouds, remains one of the main sources of uncertainty in global climate models, as it is played on scales much finer than their mesh size. A team including Paul O'Gorman (MIT), Janni Yuval and Marat Khairoutdinov replaced in the Community Atmosphere Model version 6 (CAM6) the classic Zhang-McFarlane scheme with a fully connected network of neurons, leaving radiation, turbulence and other cloudy processes intact.

Stable, and more realistic

The network learned from simulations of the gSAM model, which explicitly solves global storms, reduced to CAM6 resolution. Integrated into the model, it operates for more than a decade simulated without instability by maintaining energy and humidity, and without having to re-adjust the rest of the model. It significantly mitigates two old defects: biases on low and extreme precipitation, and the peak of rain on tropical lands, which goes from an unrealistic midday to the observed late afternoon. For the authors, this is an argument in favour of targeted parameterizations learned in next generation models.