What question did this study set out to answer?

This research aims to enhance the accuracy of radiation parameterization in the ICON climate model using machine learning techniques.

March 6, 2026Open Access

Interpreting and improving the radiation parameterization for ICON with machine learning

Key Points

This research aims to enhance the accuracy of radiation parameterization in the ICON climate model using machine learning techniques.
Developed a machine learning-based radiation emulator using a BiLSTM network.
Analyzed input variables affecting radiative predictions using Shapley Additive Explanations (SHAP).
Conducted hybrid simulations comparing the ML emulator with traditional radiation updates in ICON.
Evaluated the emulator's stability over long-term simulations and its adaptation to warming scenarios.
Improved cloud-radiation interactions using high-resolution simulations to reduce errors in heating rates.
The ML-based radiation emulator learned physical relationships consistent with radiative transfer models.
No statistically significant deviations were found between the ML model and reference simulations over one year.
The ICON-A_ML emulator showed stability over 10 years without accumulating biases.
The emulator performed four times faster than the original parameterization.
The approach reduced errors in cloud radiative effects by 4-10 times compared to traditional methods.

Abstract

With progressing climate change, we need robust climate projections supporting the development of mitigation and adaptation strategies to inform policy makers. Radiation is an important part of the climate system as it regulates the energy balance of the Earth but cannot be resolved in the spectral dimension in climate models and thus needs to be parameterized. Machine Learning (ML) has been shown to be useful to improve the performance of physical parameterizations in climate models. The main focus of this thesis is to develop an ML-based radiation scheme and to investigate possible improvements of the radiation parameterization in the ICOsahedral-Nonhydrostatic (ICON) model. In the first part of this thesis, I develop an ML-based radiation emulation for the ICON model. A BiLSTM network predicts radiative heating rates and boundary fluxes at the surface and top of the atmosphere based on the current atmospheric state including boundary conditions such as incoming solar radiation. The bidirectional part of the BiLSTM scans the vertical direction of an atmospheric column, similarly to the upward and downward fluxes in physics-based radiation schemes. To analyze what the neural network has learned, I used SHapley Additive exPlanations (SHAP) to identify which input variables determine the radiative predictions of the BiLSTM. The analysis of the Shapley values reveals that the BiLSTM learned relationships related to known physical laws: a cloud reflects incoming shortwave radiation which has a strong local effect on the heating rate, but also affects all layers below leading to less heating due to less downwelling shortwave radiation below the cloud. Additionally, the layers above the cloud are affected as well, as the reflected radiation can interact with ozone. For longwave radiation, the ambient temperature plays a significant role because it determines how much longwave radiation is emitted by the gases and clouds that are locally available. This shows that the ML-based radiation emulator learns relationships that are consistent with physical processes in radiative transfer models, which is a crucial step for its operational use in an Earth System Model (ESM). The second part of this thesis focuses on the coupling of the proposed ML-based radiation emulator to the ICON model to perform hybrid simulations and test stability. The hybrid model ICON-AML is compared with two reference setups using low and high frequent radiation updates (ICON-ALF, ICON-AHF). The comparison of 10 ensemble members of each setup with a length of one year showed that ICON-AML has no statistically significant deviations compared to the reference simulations. A long-term simulation of 10 years demonstrates that ICON-AML shows no sign of accumulated biases compared to the references ICON-AHF and ICON-ALF. This indicates that ICON-AML is stable in longer simulations while also being four times faster than the original radiation parameterization. Testing with increased prescribed sea surface temperatures showed that ICON-AML can generalize to a +4 K warmer climate. This indicates that the ML-based radiation emulator is fit-for-purpose for its use in ICON simulations. The final study of this thesis aims to improve cloud-radiation interactions in climate models utilizing high-resolution simulations. Global Storm Resolving Models (GSRM) with a horizontal resolution of 5 km resolve cloud systems better than coarse-scale models with a horizontal resolution of around 100~km. Coarse-graining high-resolution data allows to implicitly learn subgrid-scale effects related to the horizontal and vertical distribution of clouds. I develop an approach to specifically target the improvement of the cloud radiative effects on heating rates. This approach showed 4-10 times smaller errors compared to the coarse-scale physics-based radiation scheme that was evaluated on coarse-grained data. This thesis shows that the presented ML-based radiation emulator is interpretable and that it effectively mimics physical laws. The developed ML-based radiation scheme is fast and stable when coupled to a model like ICON. Additionally, it demonstrates that high-resolution simulations can be used to improve cloud radiative effects on heating rates by implicitly learning subgrid-scale effects paving the way for reducing longstanding biases in cloud radiative effects in machine learning enhanced ESMs.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper