Fast domain-aware neural network emulation of a planetary boundary layer parameterization in a numerical weather forecast model

TLDR

Parameterizations for physical processes in weather and climate models are computationally expensive, and planetary boundary layer (PBL) schemes are used to represent the diurnal evolution of turbulence, velocity, temperature, and humidity profiles that are otherwise unresolved at typical climate model scales. The study aims to train deep neural networks (DNNs) to emulate a WRF planetary boundary layer parameterization and assess their accuracy as an alternative to physics‑based schemes. Using output from the WRF climate model, the authors train domain‑aware DNNs that incorporate underlying domain structure such as nonlocal mixing to emulate the PBL parameterization. The domain‑aware DNN successfully reproduces vertical profiles of velocity, temperature, and water vapor throughout the diurnal cycle, and a single model trained at one location predicts nearby locations with correlations exceeding 0.9 compared to WRF simulations. Abstract.

Abstract

Abstract. Parameterizations for physical processes in weather and climate models are computationally expensive. We use model output from the Weather Research Forecast (WRF) climate model to train deep neural networks (DNNs) and evaluate whether trained DNNs can provide an accurate alternative to the physics-based parameterizations. Specifically, we develop an emulator using DNNs for a planetary boundary layer (PBL) parameterization in the WRF model. PBL parameterizations are used in atmospheric models to represent the diurnal variation in the formation and collapse of the atmospheric boundary layer – the lowest part of the atmosphere. The dynamics and turbulence, as well as velocity, temperature, and humidity profiles within the boundary layer are all critical for determining many of the physical processes in the atmosphere. PBL parameterizations are used to represent these processes that are usually unresolved in a typical climate model that operates at horizontal spatial scales in the tens of kilometers. We demonstrate that a domain-aware DNN, which takes account of underlying domain structure (e.g., nonlocal mixing), can successfully simulate the vertical profiles within the boundary layer of velocities, temperature, and water vapor over the entire diurnal cycle. Results also show that a single trained DNN from one location can produce predictions of wind speed, temperature, and water vapor profiles over nearby locations with similar terrain conditions with correlations higher than 0.9 when compared with the WRF simulations used as the training dataset.