Interpretable Machine Learning-Based Radiation Emulation for ICON

Kurzfassung

The radiation parameterization is one of the computationally most expensive components of Earth system models (ESMs). To reduce computational cost, radiation is often calculated on coarser spatial or temporal scales, or both, than other physical processes in ESMs, leading to uncertainties in cloud-radiation interactions and thereby in radiative temperature tendencies. One way to address this issue is to emulate the radiation parameterization using machine learning (ML), which is typically faster and has good accuracy in high-dimensional parameter spaces. This study investigates the development and interpretation of an ML-based radiation emulator using the ICOsahedral Non-hydrostatic model with the RTE+RRTMGP radiation code, which calculates radiative fluxes based on the atmospheric state and its optical properties. With a Bidirectional Long Short-Term Memory architecture, which can account for vertical bidirectional auto-correlation, we can accurately emulate shortwave and longwave heating rates with a mean absolute error of 0.045 K/d (2.77%) and 0.060 K/d (4.50%), respectively. Further, we analyze the trained neural networks using Shapley Additive exPlanations and confirm that the networks have learned physically meaningful relationships among the inputs and outputs. It is worth noting that we observe that the local temperature is used as a predictive source for the longwave heating, consistent with physical models of radiation. For shortwave heating, we find that clouds reflect radiation, leading to reduced heating below the cloud. In contrast, an architecture that is not inspired by the underlying physics, such as a multilayer perceptron, tends to rely on spurious or less physically meaningful correlations to make its predictions.