A Machine Learning-based Planetary Boundary Layer Height Scheme for ICON-A Learned from Vertically Highly Resolved Simulations

Janis Klamt, Mierk Schwabe, Axel Lauer, M. A. Giorgetta, Thorsten Mauritsen, Veronika Eyring

Preprint · 26 December 2025 · Max Planck Institute for Meteorology

In one sentence. ML boundary-layer height scheme trained on coarse-grained high-resolution ICON runs; stable 36-year AMIP coupling reduces biases in coarse-resolution climate simulations.

Abstract

In climate models, an accurate representation of the planetary boundary layer height (BLH) is critical, as it is a key input to turbulence parameterizations and directly impacts the vertical transport of momentum, heat, moisture, and tracers. Despite its importance, classical BLH schemes often exhibit systematic errors at coarse vertical resolutions, leading to biases in turbulent diffusion. Using the atmospheric component of the ICON model, we investigate the sensitivity of three classical BLH schemes to vertical model resolution. We identify biases and numerical artifacts in BLH distributions when using the standard ICON-A configuration with coarse vertical resolution. Simulations with substantially higher vertical resolution show improved agreement with ERA5 reanalysis data. However, such high-resolution configurations are computationally expensive and not feasible for long-term climate simulations. To address this limitation, we develop a machine learning (ML)-based BLH scheme trained on vertically high-resolution ICON simulations that are coarse-grained to the standard climate model configuration. The ML-based scheme achieves a coefficient of determination exceeding $R^2=0.99$ relative to high-resolution BLH estimates. It effectively reduces the biases of classical schemes in low-resolution configurations, especially improving the diurnal cycle over complex terrain. When coupled online to ICON, the ML scheme enables stable 36-year AMIP simulations without further tuning and leads to changes in surface climatology and general circulation. It improves both the accuracy and statistical distribution of simulated BLHs, yielding a more consistent representation of boundary layer processes in climate simulations. Moreover, relative to observational datasets, the ML scheme shows slight improvements in cloud cover and precipitation.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.