Online Test of a Neural Network Deep Convection Parameterization in ARP-GEM1

TL;DR

This study has successfully replaced ARP-GEM1's deep convection scheme with a neural network emulator, and the evaluation of averaged fields showed good agreement with output from an ARP-GEM1 simulation using the physics-based deep convection scheme.

Abstract

In this study, we present the integration of a neural network-based parameterization into the global atmospheric model ARP-GEM1, leveraging the Python interface of the OASIS coupler. This approach facilitates the exchange of fields between the Fortran-based ARP-GEM1 model and a Python component responsible for neural network inference. As a proof-of-concept experiment, we trained a neural network to emulate the deep convection parameterization of ARP-GEM1. Using the flexible Fortran/Python interface, we have successfully replaced ARP-GEM1's deep convection scheme with a neural network emulator. To assess the performance of the neural network deep convection scheme, we have run a 5-years ARP-GEM1 simulation using the neural network emulator. The evaluation of averaged fields showed good agreement with output from an ARP-GEM1 simulation using the physics-based deep convection scheme. The Python component was deployed on a separate partition from the general circulation model, using GPUs to increase inference speed of the neural network.

Figures (5)

• Figure 1: Field exchange between ARP-GEM1 and the Python-based NN inference component. In the online experiment, ARP-GEM1 computes the dynamics tendencies first. During the physical tendencies computation (including subgrid-scale parameterizations), ARP-GEM1 sends the necessary NN input fields to the Python component. The Python process performs the NN inference and sends back the NN output tendencies for deep convection. These tendencies are incorporated into the total tendencies in ARP-GEM1, which are then integrated and used by the dynamics at the next timestep.
• Figure 2: Offline evaluation of the NN. $R^2$ scores were computed using the test sample. Global $R^2$ score of the model is 0.33.
• Figure 3: Zonal average values of tendencies computed on the offline validation sample, for the offline test period (01 Jan 2006 to 31 Dec 2006). Data has been interpolated from model levels to pressure levels. The left column shows the target variables from the validation sample, while the right column presents the NN inference results on the same validation sample. The top row represents dry static energy tendencies, $\partial_t s$, and the bottom row displays specific humidity tendencies, $\partial_t q$, both from deep convection processes.
• Figure 4: CERES observations regridded to TCo179 resolution and anomalies of outgoing longwave radiation (rlut, top half) and net shortwave radiation (rst, bottom half). The average values are computed from simulations covering a 5-year period from 01 Jan 2006 to 31 Dec 2010, from different ARP-GEM1 simulations (see Section \ref{['sec:online']}).
• Figure 5: MSWEP observations regridded to TCo179 resolution and total precipitation anomalies. The average values are computed from simulations covering a 5-year period from 01 Jan 2006 to 31 Dec 2010, from different ARP-GEM1 simulations (see Section \ref{['sec:online']}).