Improvement of a Neural Network Convection Scheme by Including Triggering and Evaluation in Present and Future Climates

In this study, we improve a neural network (NN) parameterization of deep convection in the global atmosphere model ARP-GEM. To account for the sporadic nature of convection, we develop an NN parameterization that includes a triggering mechanism capable of detecting whether deep convection is active within a grid cell. This new data-driven parameterization outperforms the existing NN parameterization in present climate when replacing the original deep convection scheme of ARP-GEM. Online simulations with the NN parameterization run without stability issues. Then, this NN parameterization is evaluated online in a warmer climate. We also confirm that using relative humidity instead of the specific total humidity as input for the NN (trained with present data) tends to improve the performance and generalization in warmer climate. Finally, we perform the training of the NN parameterization with data from a warmer climate, and this configuration gets similar results when used in simulations in present or warmer climates. Significance Statement This paper introduces a data-driven parameterization that significantly improves upon the method described in Balogh et al. Incorporating physical knowledge in the data-driven parameterization leads to reduced biases in the simulation using the data-driven parameterization and enhances both online performance and stability, including in climates not encountered during the data-driven parameterization’s training phase.