A Lightweight Neural Network for Accelerating Radiative Transfer Modeling in WRF
Erick Fredj 1 2 , Iggy Segev Gal 2 , Noam Lavi 2 , Shahar Belkar 2 , Mark Wasserman 2 , Ding Zhaohui 2 , Yann Delorme 2 · 2025
Read A Lightweight Neural Network for Accelerating Radiative Transfer Modeling in WRF by Erick Fredj 1 2 , Iggy Segev Gal 2 , Noam Lavi 2 , Shahar Belkar 2 , Mark Wasserman 2 , Ding Zhaohui 2 , Yann Delorme 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-14542.txt.
Introduction
Erick Fredj and co-authors speed up weather math in a 2025 paper. Weather models must work out how heat and light move through air, a step that eats much computer time. The classic methods are exact but slow. Neural networks can copy the step at a small part of the cost, yet their accuracy is often not tested well for rare, harsh weather.
The team builds a light network for the WRF weather model, a widely used forecast system. They train it to stand in for the slow radiation step. Then they check it hard on an extreme weather case, where training data is thin. The paper aims to show the swap keeps forecast quality while cutting the time the step takes.
Catalog record
| CREATOR | Erick Fredj 1 2 , Iggy Segev Gal 2 , Noam Lavi 2 , Shahar Belkar 2 , Mark Wasserman 2 , Ding Zhaohui 2 , Yann Delorme 2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-14542
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Lightweight Neural Network for Accelerating Radiative Transfer Modeling in WRF * Thanks: Erick Fredj 1 2 , Iggy Segev Gal 2 , Noam Lavi 2 , Shahar Belkar 2 , Mark Wasserman 2 , Ding Zhaohui 2 , Yann Delorme 2 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. I Introduction The computation of the Rapid Radiative Transfer Model for General Circulation Models (RRTMG) within the Weather Research and Forecasting (WRF) model constitutes a substantial portion of the total computational effort, ranging from up to 20% on clear sky days to as much as 60% on cloudy days. Although the specific percentage varies depending on the model configuration, radiative transfer calculations are particularly time-consuming components of WRF models (e.g., [ 1 ] , [ 2 ] ; [ 3 ] ). In both climate modeling and numerical weather prediction (NWP), there is an inherent trade-off between the accuracy and computational efficiency of radiative transfer calculations. Highly accurate methods, such as line-by-line procedures, could theoretically compute radiative fluxes at every grid point and every time step. However, performing radiative transfer calculations at this level of detail would require as much or more CPU time than all other model components combined, including model dynamics and other physical parametrization [ 2 ] . To manage this computational load, various simplifications are often implemented. One common approach is the correlated-k method (Lacis and Oinas, 1991), which reduces the integration over wavelength by grouping wavelengths with similar absorption coefficients (k terms). This significantly decreases the number of monochromatic radiative transfer calculations needed. The number of k terms can be adjusted to balance accuracy and efficiency for a specific application; however, the correlated-k methods are not efficient enough to allow calculations at every grid point and time step. To further reduce computational costs, radiative calculations are typically performed at reduced temporal and/or spatial resolutions. For example, in climate and global forecast models at the National Centers for Environmental Prediction (NCEP) and the United Kingdom Met Office (UKMO), radiative calculations are performed every 1 to 3 hours [ 3 ] . This means that significant changes in radiative profiles, due to factors like cloud cover and the angle of incident solar radiation, might not be captured between these calculations. At the European Center for Medium-Range Weather Forecasts (ECMWF), radiative calculations are conducted on a coarser grid and then interpolated to a finer grid, speeding up the calculations [ 1 ] , [ 2 ] . Similarly, the Canadian operational Global Environmental Multi-scale model uses a reduced vertical resolution approach, calculating full radiation at alternate vertical levels and interpolating for intermediate levels (e.g., [ 4 ] , [ 5 ] ). These strategies reduce the horizontal or vertical variability in radiation fields and can compromise the accuracy and consistency of the model’s radiation calculations, potentially affecting climate simulations and weather predictions. The trade-off between computational speed and accuracy in radiation calculations has led to the development of alternative numerical algorithms designed to enhance computational efficiency while maintaining accuracy. Two techniques proposed to improve the temporal and spatial resolution of radiation calculations involve enhanced interpolation methods for radiative calculations from a coarse to a fine grid [ 2 ] and improved radiative calculations between full-time steps ( [ 6 ] ; [ 3 ] ). However, these approaches can introduce significant errors over time due to accumulated discrepancies in interactions with other dynamic and physical processes ( [ 7 ] ). Given the need to balance speed and accuracy, data-driven radiation emulators based on neural networks (NNs) have emerged as promising alternatives ( [ 8 ] ; [ 9 ] ; [ 10 ] ), offering significant improvements in computational speed while maintaining reasonable accuracy. Chevallier et al. ( [ 11 ] , [ 12 ] ) pioneered NN-based longwave (LW) radiation emulation for ECMWF models. Since then, NN-based LW and shortwave (SW) emulators have been developed for various models, including the Community Atmosphere Model (CAM), the Climate Forecast System (CFS), and the Super-Parameterized Energy Exascale Earth System Model (SP-E3SM) ( [ 13 ] ; [ 14 ] ; [ 15 ] ; [ 16 ] ; [ 17 ] ). Krasnopolsky et al. [ 16 ] reported impressive results for an NN-based emulator of the Rapid Radiative Transfer Model for General Circulation Models (RRTMG; [ 18 ] ; [ 19 ] ), achieving computational speed improvements of 16-60 times while preserving long-term (17-year) stability. In Belochitski et al. [ 13 ] , the NN-based radiation emulator outperformed those based on Classification and Regression Trees (CART). Recently, Pal et al. [ 17 ] achieved a tenfold speed increase and 90-95 % accuracy using a deep neural network (DNN), despite the greater computational demands of DNNs. Various emulators have also been developed for idealized frameworks (e.g., [ 20 ] ; [ 21 ] ) and specific processes such as convection [ 22 ] , the planetary boundary layer [ 23 ] , and dynamics. While advanced deep-learning techniques have been applied to post-processing in weather and climate models, emulating dynamic and physical processes within numerical models remains a challenging task. In this study, we develop a new lightweight neural-network-based model for radiative heat transfer, and investigate its performance within the Weather Research and Forecasting (WRF) model. We follow the approach of Krasnopolsky et al. [ 20 ] to develop a neural network (NN) emulator for both longwave and shortwave radiation physics formulations. The accuracy of the model is demonstrated for the challenging forecast of Typhoon Muifa, which made landfall near China’s largest metropolitan area and several major shipping ports on September 14, 2022. The following sections provide a detailed description of the methodology. Section II details specifics about the modeling approach and data used to train the neural network. A qualitative and quantitative evaluation of the accuracy of the model, along with performance analysis, is given in Section III . Finally, a summary and prospect for future research are presented in Section IV . II Modeling Approach II-A Neural Network Emulator for RRTMG Model This study proposes a light-weight neural network emulator for the Rapid Radiative Transfer Model for GCMs (RRTMG), composed of a single hidden layer. The proposed neural network includes 178 input variables for the clear-sky model, and 222 input variables for the cloudy-sky model. The output variables consist of 47 longwave (LW) and shortwave (SW) components, as shown in Table I . Inputs (178,222) Vertical Pressure 44 Vertical Temperature 44 Vertical Ozone 44 Vertical Cloud fraction (cloudy) 44