Application of neural networks for improving surface layer turbulent exchange parameterizations in general circulation models

Turbulent heat and momentum transfer in atmospheric surface layer plays a crucial role in determining the global and regional climate. They have a significant impact on various processes, such as the interaction between the atmosphere and the underlying surface, as well as on atmospheric circulation.

Methods based on the Monin-Obukhov similarity theory (MOST) and its generalizations are widely used in numerical weather and climate models. However, their limitations become especially evident in stable conditions, such as those found in the polar regions.  MOST assumes the existence of constant flux layer, where the dimensionless gradients of wind speed, temperature, and humidity depend on the height above the surface relative to the Obukhov length scale. This assumption is violated in conditions of a thin stably stratified boundary layer (about 10-100 meters). In addition, MOST does not take into account the non-gradient transport characteristics of the daytime convective boundary layer. It also does not consider the influence of topography and thermal heterogeneity, which can lead to systematic errors in the estimation of turbulent fluxes and increase uncertainty in climate predictions.

This study discusses the use of a neural network-based algorithm that can take into account complex, nonlinear relationships between meteorological parameters. This algorithm may compensate for some of the limitations of known semi-empirical approaches. The model was trained on extensive datasets obtained from field experiments, such as MOSAiC and SHEBA. This allows the model to capture not only the general physical relationships between input parameters and fluxes, but also to accurately reproduce the amplitude of turbulent fluxes, even under strongly stable conditions. The possibility of training a model using the output data from a single column atmospheric boundary layer model with fine resolution tuned to large-eddy simulation data is also being explored. This allows us to consider the influence of the height of the boundary layer, thereby expanding the applicability of the algorithm to climate models with coarse vertical resolution. The results obtained demonstrate a significant improvement in the accuracy of flux estimates compared to the surface layer parameterization used in the INMCM¹ Earth system model. This suggests the potential for integrating machine learning into surface layer models, which in turn may contribute to improving weather and climate forecast.

¹Link to INMCM: https://link.springer.com/article/10.1007/s00382-017-3539-7

Volodin, E.M., Mortikov, E.V., Kostrykin, S.V. et al. Simulation of the present-day climate with the climate model INMCM5. Clim Dyn 49, 3715–3734 (2017). https://doi.org/10.1007/s00382-017-3539-7