Abstract

AbstractAtmospheric radiation is the main driver of weather and climate, yet due to a complicated absorption spectrum, the precise treatment of radiative transfer in numerical weather and climate models is computationally unfeasible. Radiation parameterizations need to maximize computational efficiency as well as accuracy, and for predicting the future climate many greenhouse gases need to be included. In this work, neural networks (NNs) were developed to replace the gas optics computations in a modern radiation scheme (RTE+RRTMGP) by using carefully constructed models and training data. The NNs, implemented in Fortran and utilizing BLAS for batched inference, are faster by a factor of 1–6, depending on the software and hardware platforms. We combined the accelerated gas optics with a refactored radiative transfer solver, resulting in clear‐sky longwave (shortwave) fluxes being 3.5 (1.8) faster to compute on an Intel platform. The accuracy, evaluated with benchmark line‐by‐line computations across a large range of atmospheric conditions, is very similar to the original scheme with errors in heating rates and top‐of‐atmosphere radiative forcings typically below 0.1 K day−1 and 0.5 W m−2, respectively. These results show that targeted machine learning, code restructuring techniques, and the use of numerical libraries can yield material gains in efficiency while retaining accuracy.

Highlights

  • Atmospheric radiation is the fundamental energy source which drives weather and climate

  • Atmospheric radiation is the main driver of weather and climate, yet due to a complicated absorption spectrum, the precise treatment of radiative transfer in numerical weather and climate models is computationally unfeasible

  • For a machine-learned radiation scheme to perform well in simulations of future climate, for example, it is important that future concentrations of greenhouse gas (GHG) and warmer and moister atmospheres are readily sampled during training, as neural networks (NNs) are unlikely to extrapolate beyond the trained input space with much skill

Read more

Summary

Introduction

Atmospheric radiation is the fundamental energy source which drives weather and climate. Code optimization can play an important role: Current NWP and climate models often have low arithmetic intensity and underutilize the computational power of modern supercomputers, but code restructuring techniques can significantly improve performance by alleviating memory bottlenecks and improving vectorization (Michalakes et al, 2016) Another interesting alternative is machine learning (ML), which is currently a popular research topic in the context of physical modeling, as it has the potential to reduce key sources of uncertainty in dynamical models. NNs have previously been used to emulate the entire radiation scheme in a dynamical model, with the outputs being the radiative fluxes and heating rates for all layers in an atmospheric column (Krasnopolsky et al, 2010; Pal et al, 2019) This approach has yielded considerable speed-ups of one (Pal et al, 2019) or several (Krasnopolsky et al, 2010) orders of magnitude compared to the original scheme. The speed and accuracy of the new gas optics code, which is coupled to a refactored solver, is evaluated against the original scheme using benchmark LBL computations

RRTMGP
Background
Model Design
Model Training and Tuning
Implementation and Code Optimization
Accuracy
Fast and Flexible Models
Findings
Data Availability Statement
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call