Improvement of a neural network convection scheme by including triggering and evaluation in present and future climates

2025

Read Improvement of a neural network convection scheme by including triggering and evaluation in present and future climates (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-05074.txt.

Introduction

Hugo Germain and co-authors improve a neural net scheme for deep convection inside a global weather model. Storm clouds are sporadic, so a scheme must first tell whether deep convection is on or off in each grid cell.

Their new scheme adds that trigger step to the net. Swapped into the ARP-GEM model, it beats the prior net scheme in the current climate and runs online with no stability trouble. Trials in a warmer climate also hold up, and the choice of humidity input helps the scheme carry to new conditions.

Catalog record

YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-05074

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Improvement of a neural network convection scheme by including triggering and evaluation in present and future climates Abstract In this study, we improve a neural network (NN) parameterization of deep convection in the global atmosphere model ARP-GEM. To take into account the sporadic nature of convection, we develop a NN parameterization that includes a triggering mechanism that can detect whether deep convection is active or not 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 confirm that using relative humidity instead of the specific total humidity as input for the NN (trained with present data) improves 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 get similar results when used in simulations in present or warmer climates. † † authors: Hugo Germain \aff a, Blanka Balogh \aff a, Olivier Geoffroy \aff a, David Saint-Martin \aff a \statement This paper introduces a data-driven parameterization that significantly improves upon the method described in 3 . Two key advancements are presented, leading to reduced biases in the simulation using the data-driven parameterization. First, a triggering mechanism is incorporated in the data-driven parameterization, which effectively mitigates biases. Second, the replacement of absolute humidity with relative humidity as an input enhances both online performance and stability, including in climates not encountered during the data-driven parameterization’s training phase. 1 Introduction Parameterizations of atmospheric moist processes are the main source of biases in current climate models ( 25 ; 26 ; 34 ) . The use of Machine Learning (ML) techniques, especially Neural Networks (NNs), to develop data-driven parameterizations is a promising approach to significantly improve the accuracy of climate models ( 15 ) . During the past decade, data-driven approaches were widely used to develop parameterizations for climate models. NNs were used to produce accurate, yet numerically affordable radiative transfer schemes ( 12 ; 23 ; 36 , e.g.,) , cloud microphysics ( 32 ; 31 ) or convection ( 9 ; 3 , e.g.,) . They have been used to emulate subgrid-scale parameterizations from aggregated high-resolution simulations ( 43 ; 42 , e.g.,) or from a super-parameterized model ( 15 ; 30 , e.g.,) . Until recent years, only a few simulations using data-driven parameterizations were carried out, as a substitute for traditional physical ones. However, significant technical advancements in integrating NNs into Fortran-based models have now made it easier to perform online tests of data-driven parameterizations. 8 conducted an online evaluation of a data-driven unified parameterization in a single column model, extended to a full General Circulation Model (GCM) in 10 and 9 , with a focus on online stability of the data-driven scheme. 37 also used NNs trained using SPCAM data to represent the subgrid-scale processes in the atmospheric model CAM5 ( 27 ) . The NN parameterization described in 38 was based on the output of a global storm-resolving simulation using GFDL X-SHiELD ( 20 ) to represent of heating and moistening rates in the Global Forecast System ( 44 , GFS,) . ClimSim Online ( 41 ) implemented Pytorch-Fortran ( 1 ) to conduct an experiment with a data-driven parameterization based on the ClimSim dataset 40 in the E3SM model ( 29 ) . Using FTorch ( 2 ) in the ICON-A model ( 18 ) , a several data-driven parameterizations were tested online , such as deep convection ( 21 ) (stable online for 180 days) and radiative transfer 19 . 3 (hereafter, B25) used the OASIS-coupler’s Fortran/Python interface ( 13 ) to replace heating and moistening tendencies of a deep convection parameterization by NNs in the ARP-GEM global atmosphere model, version 1 ( 17 ) . To evaluate the online performance of the NN-based deep convection parameterization, B25 carried out a 30-year simulation using ARP-GEM. The simulation produced realistic physical fields for most variables. However, it exhibited some biases, particularly in high cloud cover and over the polar regions. In this paper, we aim to present two major improvements to the data-driven deep convection parameterization introduced in B25, addressing the biases we have identified using the ARP-GEM atmosphere model, version 2 ( 16 ) . The first improvement involves using a triggering mechanism. Second, following the suggestion in 6 , we replace absolute humidity by relative humidity (RH) to improve the generalizability of the data-driven scheme. The following manuscript is organized as follows. The first section describes the data-driven parameterization, including the data-driven triggering mechanism, and its performance both offline and online. The second section extends the online evaluation of the data-driven parameterization by testing its generalizability in a different climate. 2 A ML-parameterization with triggering mechanism 2.1 Model description We use the global efficient and multi-resolution atmosphere model ARP-GEM version 2 ( 16 ) with minor modifications described below. The model configuration is the same as in B25 with a horizontal resolution of 55 km and 50 hybrid coordinate vertical levels, extending from the surface up to 2 hPa. The model time step is set to \Delta t = 900 s. Some modifications have been made to the model since the study of B25, hence our results are not directly comparable with B25. B25 use ARP-GEM version 1. Here we use ARP-GEM version 2. Differences concern mainly the shallow convection scheme and model tuning. The triggering mechanism has also slightly been revised with a different formulation of entrainment in the triggering test parcel. These differences are described in detail in 16 . The deep convection parameterization of ARP-GEM is based on 35 revised by 4 ; 5 ; 14 ; 17 and 16 and will be referred to as the Tiedtke-Bechtold scheme thereafter. Entrainment and detrainment rates are higher than in B25, with the coefficients \epsilon_{up} and \delta_{up} , as defined in 14 , set to 2.0 \cdot 10 -3 m -1 and 0.8 \cdot 10 -4 m -1 , respectively, instead of 1.8 \cdot 10 -3 m -1 and 0.75 \cdot 10 -4 m -1 in B25. Additionally, the intensity of shallow convection is reduced by a factor of three. Finally, for simplicity, the shallow cloud cover is set to zero in the model version used here. The differences in model physics, particularly those related to deep convection, explain differences in results when replicating B25, as mentioned in Section 2 . 2.4 . 2.2 A NN parameterization with triggering The Tiedtke-Bechtold scheme computes tendencies of dry static energy \partial_{t}s , specific humidity \partial_{t}q and zonal and meridional winds. For simplicity, we only emulate the thermodynamical tendencies ( \partial_{t}s and \partial_{t}q ) giving they constitute the main tendencies of the model. The momentum tendencies are still computed by the Tiedtke-Bechtold parameterization. Because both thermodynamic and momentum tendencies are computed following the same framework, under the assumption that thermodynamical tendencies are well represented by the NN, the representation of momentum tendencies should be straightforward. Reproducing non gaussian processes can be challenging for NNs ( 33 ) and it may produces artificial signals. This is particularly true

Read the complete text