A Lightweight Neural Network for Accelerating Radiative Transfer Modeling in WRF

TL;DR

This work tackles the high computational cost of radiative transfer in WRF by introducing a lightweight neural-network emulator for the RRTMG radiation scheme. The model uses a single hidden layer with region-specific inputs, split into eight sub-models to cover LW/SW and land/ocean conditions, and is trained on large WRF-generated datasets. Validations on Typhoon Muifa demonstrate high fidelity to the full solver, with substantial speedups (up to ~2x faster overall) and strong spatial-temporal agreement, including near-perfect Pearson correlations for key flux components. Temporal transfer learning further enables year-round applicability, enabling efficient, accurate forecasts for extreme events and routine weather periods alike.

Abstract

Radiative transfer calculations in weather and climate models are notoriously complex and computationally intensive, which poses significant challenges. Traditional methods, while accurate, can be prohibitively slow, necessitating the development of more efficient alternatives. Recently, empirical emulators based on neural networks (NN) have been proposed as a solution to this problem. These emulators aim to replicate the radiation parametrization used in the models, at a fraction of the computational cost. However, a common issue with these emulators is that their accuracy has often been insufficiently evaluated, especially for extreme events for which the amount of training data is sparse. The current study proposes such a model for accelerating radiative heat transfer modeling in WRF, and validates the accuracy of the approach for an extreme weather scenario.

Figures (11)

• Figure 1: Convergence of the AI models for clear and cloudy conditions, longwave (LW) and short wave (SW).
• Figure 2: Understanding the Machine Learning (ML) workflow model. Learning curves for real cases. The results for the sequential neural network (SNN) were derived using the optimal settings identified through subsequent analyses. The optimal epoch and the normalized root mean square error (RMSE) for all outputs are provided in parentheses.
• Figure 3: Histogram showing the distribution of the cloud fraction center of mass for 412,704 training samples of a cloudy model. Noticeable declines in sample counts at low and high center of mass values underscore the necessity of dividing the data into folders as described to ensure that models are trained with a diverse range of cloud scenarios, enhancing their predictive accuracy across different atmospheric conditions
• Figure 4: Domain configuration for Zhejiang Province: - Reference Coordinates: Latitude 29, Longitude 123.5 - Horizontal Grid Projection: Lambert Conformal Conic - Grid dimensions: 255x225 and 370x370 points in the x and y directions - Grid spacing: 9 km and 3 km in both x and y directions.
• Figure 5: Simulation duration of Typhoon Muifa by the WRF-RRTMG and WRF-NN models. The WRF-RRTMG track is represented by black dots, and the WRF-NN track is depicted by red dots.