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

Abstract In this study, we integrate a neural network–based parameterization into the global atmospheric model ARP-GEM1 using the Python interface of the OASIS coupler. This setup enables the exchange of fields between the FORTRAN-based ARP-GEM1 model and a Python component implementing the neural network inference. The Python component was deployed on a separate partition from the general circulation model, using graphics processing units (GPUs). As a proof of concept, we trained a neural network to emulate ARP-GEM1’s deep convection parameterization. Leveraging the flexible FORTRAN–Python interface, we successfully replaced ARP-GEM1’s deep convection scheme with the neural network emulator. To evaluate its online performance, we realized a 30-yr ARP-GEM1 simulation using the neural network for deep convection. The evaluation of the averaged fields showed good agreement with the output of an ARP-GEM1 simulation using the physics-based deep convection scheme. Significance Statement In this study, we have successfully coupled a FORTRAN-based atmospheric general circulation model and a Python-based component implementing a neural network. Our coupling approach provides flexibility by allowing the atmospheric model to run on central processing units and the Python component to run concurrently on graphical processing units. As a simple test case of our implementation, we replaced part of the atmospheric model with a Python-based neural network. The resulting hybrid model was used to realize a 30-yr simulation that produced realistic climatologies of physical fields.