From stable online coupling to decade-long climate simulations: A machine learning parameterization for cloud microphysics in ICON

2026

Read From stable online coupling to decade-long climate simulations: A machine learning parameterization for cloud microphysics in ICON (2026) 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-2606-23829.txt.

Introduction

Sarauer and colleagues run decade long climate simulations with a machine learned treatment of cloud droplet physics inside a major Earth system model, in a 2026 paper on climate modelling. Cloud microphysics, the birth and growth of droplets and ice crystals, is a stubborn source of error in climate models. The team trains a neural scheme on a global simulation fine enough to resolve convection directly, then couples it into the ICON model at coarse resolution.

The scheme uses a classifier to find active cells and a regressor to predict the changes. Physical guardrails, such as never letting cloud mass go negative, prove essential. The coupled model runs stably for a decade and matches the observed climate about as well as the classical scheme. It also removes two tuning knobs.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE83 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.23829

Browse the full library · About the Tower of Babel

Full text

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

From stable online coupling to decade-long climate simulations: A machine learning parameterization for cloud microphysics in ICON Abstract The representation of cloud microphysics and its nonlinear character and scale-dependence is a remaining source of uncertainty in Earth system models (ESMs). Here, we develop and couple online a machine learning (ML)-based cloud microphysics parameterization with the Icosahedral non-hydrostatic modeling framework (ICON). The primary challenge is achieving numerically stable, long-term online coupling when transitioning from training with km-scale data to application in coarse-scale simulations, where the coupled system encounters atmospheric states and feedbacks not seen during training. The training data is obtained from a global convection-permitting ICON simulation at 5 kilometer resolution, providing explicitly resolved convective dynamics unavailable at coarse scales. The ML microphysics scheme uses a two-stage design: a classifier to identify active grid cells and a regressor to predict cloud microphysical tendencies. Physical constraints such as enforcing mass positivity and overshoot prevention prove essential for numerical stability in the coupled system. We demonstrate that achieving stable online coupling requires enforcing physical constraints and careful dataset curation, and that strong offline performance alone is insufficient. The coupled model maintains numerical stability over decade-long simulations with a performance in reproducing the observed climate comparable to the classical graupel scheme. The ML-based scheme eliminates two microphysics-specific tuning parameters of the classical graupel scheme, directly reducing the scope for error compensation by tuning in future hybrid models, though systematic improvements in long-term mean-state biases are not yet realized. This study demonstrates that stable, decade-long climate simulations with an ML-based cloud microphysics scheme trained on convection-permitting data are feasible, providing a foundation for future hybrid ESMs. † † journal: Machine Learning: Earth † † authors: Ellen Sarauer, Mierk Schwabe, Philipp Weiss, Axel Lauer, Philip Stier, Veronika Eyring † † corresponding: Ellen Sarauer, [email protected] Deutsches Zentrum für Luft- und Raumfahrt e.V., Institut für Physik der Atmosphäre, Oberpfaffenhofen, Germany Department of Physics, Atmospheric, Oceanic and Planetary Physics, University of Oxford, Oxford, UK European Centre for Medium-Range Weather Forecasts, Bonn, Germany Institute of Environmental Physics (IUP), University of Bremen, Bremen, Germany keypoints We present stable, decade-long climate simulations of an ML-based cloud microphysics scheme coupled to the Earth system model ICON, trained on global convection-permitting simulations to capture subgrid-scale dynamics-microphysics interactions. With the ML-based scheme, we can eliminate several tuning parameters of the classical graupel parameterization and potentially reduce the scope for compensating errors in future hybrid models; although mean-state biases are not yet reduced. We document a path to decade-long online stability, showing that physical constraints and iterative testing with the coupled model are key and confirm that strong offline performance alone is not sufficient 1 Introduction Uncertainties in cloud microphysics processes, particularly involving mixed-phase and ice clouds, and the challenge of modeling many nonlinear processes on temporal and spatial scales covering several orders of magnitude remain a major source of error in climate projections ( 41 ; 35 ) . The cloud microphysics parameterization is a core component of Earth System Models (ESMs), as they describe the formation, growth, and removal of hydrometeors and associated phase transitions. These processes have a strong influence on climate through their impact on cloud lifetime, radiative effects, and the redistribution of water and energy throughout the atmosphere. In ESMs, parameterizations are used to represent the statistical effects of unresolved processes at the grid scale. ML has been successfully used to improve ESMs by replacing or augmenting traditional parameterizations that are based on empirical relationships and physical assumptions ( 22 ; 21 ; 26 ; 8 ; 15 ) . ML models are typically trained on short, high-resolution simulations or observational data and then coupled to coarse-resolution climate models, with the aim of reducing long-standing biases in ESMs ( 17 ; 14 ) . Over recent years, learning cloud microphysics parameterizations with ML has become an active area of research ( 19 ; 45 ; 54 ; 55 ; 33 ) . A first line of work has focused on emulating cloud microphysics parameterizations within the same model setup used to generate the training data ( 19 ; 45 ) . These approaches primarily aim to improve computational efficiency by accelerating ESMs while closely reproducing the output of existing parameterizations, rather than improving the physical representation of microphysics processes. Beyond a full-scheme emulation, some studies have targeted specific cloud microphysical processes such as ice nucleation ( 18 ; 34 ) , offering a more modular path toward improving process-level realism. Other studies have trained ML models on more detailed microphysics representations using superdroplet box models ( 54 ; 56 ) . These approaches seek to enhance microphysics realism and reduce biases in coarse-resolution ESMs by using detailed, high-resolution process simulations that are computationally too expensive for application in classical parameterizations for ESMs. However, the transition from superdroplet box models to global climate simulations introduces challenges related to scale separation, boundary conditions, and interactions with dynamics. In addition, the computational characteristics of ML-based superdroplet parameterizations differ substantially from those of classical graupel schemes typically used in ESMs, adding technical complexity when implementing such models in global climate models such as ICON to simulate the highly nonlinear climate system. In our previous work ( 50 ) , we developed an ML microphysics scheme trained on convection-permitting ICON simulations (5-km horizontal resolution), then coarse-grained to the 80-km target resolution. This design allowed the ML model to capture nonlinear interactions between subgrid-scale dynamics and microphysics, as explained in 50 . The offline evaluation demonstrated strong predictive skill ( R^{2} > 0.74 ) and physically consistent behavior confirmed through explainable AI analysis. Moreover, this ML-based scheme learns an entire microphysics parameterization including ice processes, which remain a particular challenge in atmospheric modeling ( 41 ) , though its behavior and robustness in an Earth system model remained untested. Previous studies have demonstrated that strong offline performance of ML microphysics schemes does not necessarily guarantee stability once coupled online to a host model, emphasizing the importance of implementing constraints on the ML models ( 26 ; 8 ; 31 ; 60 ) . In aquaplanet experiments, ML microphysics schemes often developed instabilities, whereas random forest models produced bounded predictions and remained stable ( 23 ; 11 ) . For microphysics, the SuperdropNet emulator of warm-rain collision-coalescence has been successfully coupled to ICON in idealized bubble experiments, demonstrating stable behavior while preserving physical consistency ( 4 ) . However, despite growing interest, relatively few studies have demonstrated long-term online coupling of ML-based microphysics in realistic climate simulations. 45 demonstrated stable year-long coupling of an ML emulator of an operational microphysics scheme in a global atmospheric model, highlighting that extended online stability is achievable beyond idealized configurations. To date, no study has demonstrated decade-long stable online coupling of a complete ML microphysics scheme including ice processes trained on convection-permitting simulations. Ensuring stability and physical consistency of ML-based microphysics across climate-relevant timescales remains unresolved, particularly for parameterizations trained on convection-permitting simulations ( 35 ) . To our knowledge, this study presents the first decade-long stable online coupling of a complete ML cloud microphysics scheme trained on convection-permitting (5 km) data, performing scale transfer to a typical climate model resolution (80 km), addressing a critical and largely unsolved challenge in hybrid

Read the complete text