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

TL;DR

The paper addresses biases in existing data-driven deep convection parameterizations within the ARP-GEM model and presents two key improvements: a triggering mechanism to activate convection only when physically warranted, and replacing specific total humidity with relative humidity as an input to improve generalization. It introduces a two-NN system (MLP Classifier and MLP Predictor) that determines activation probability $p$ and generates tendencies when $p> abla\alpha$, with a balanced training dataset. Offline and online evaluations show that the triggering-enabled NN (NN-t0.5) reduces high-cloud and OLR biases and yields precipitation fields closer to observations than prior NN approaches, while remaining stable in online tests. The study also demonstrates that RH-based inputs enhance performance in a warmer (+4K) climate, and training in warmer climates generalizes well, suggesting robustness and practical potential for climate-model improvements using data-driven parameterizations.

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.

Figures (8)

• Figure 1: Scheme of the parameterization with two NNs. Inputs : profiles (50 levels) of temperature ($\mathbf{T}$), specific total humidity ($\mathbf{q}_t$) and of vertical velocity ($\mathbf{w}$) and 4 scalar fields : land-sea mask (LSM), surface pressure ($P_s$), Latent heat flux (LHF) and Sensible heat flux (SHF). Outputs : profiles (42 levels) of dry static energy tendencies ($\partial_t\mathbf{s}$) and specific humidity tendencies ($\partial_t\mathbf{q}$). $\alpha$ is a threshold to be tuned.
• Figure 2: Receiver Operating Characteristic (ROC) curve of the MLP classifier.
• Figure 3: RMSE profiles of a) dry static energy and b) humidity tendencies for the NN-nt parameterization, in orange and the NN-t0.5 parameterization, in green, computed on the validation dataset (year 2006) and interpolated to sixteen pressure levels.
• Figure 4: Differences from the zonal mean reference of a) dry static energy and b) humidity tendencies for the parameterization with 1 NN (NN-nt) and c) dry static energy and d) moisture tendencies for the NN-t0.5 parameterization computed on the validation dataset (year 2006).
• Figure 5: Anomaly with respect to an ARP-GEM reference simulation (with Tietdke-Bechtold scheme) for the simulation with the NN-nt parameterization (a) high clouds, c) OLR, e) precipitations) and for the simulation with the NN-t0.5 parameterization (b) high clouds, d) OLR, f) precipitations).