ClimART: A Benchmark Dataset for Emulating Atmospheric Radiative Transfer in Weather and Climate Models

Physics-Incorporated Framework for Emulating Atmospheric Radiative Transfer and the Related Network Study

“…In this way, the influence of the number of the parameters can be ruled out, and the influence of the network structures on the radiative transfer modeling can be examined more clearly. As the fully connected networks and convolutional-based NN are studied by many researchers before (Krasnopolsky et al, 2010;Liu et al, 2020;Cachay et al, 2021;Lagerquist et al, 2021;Ukkonen, 2022), the details are described in Text S3 in the supporting information.…”

“…However, the above methods and results were established using either incomprehensive datasets or non-common radiation schemes. Cachay et al (2021) introduced ClimART, a dataset for applications of ML in radiative transfer problems. The ClimART dataset only took into account the pristine sky (no aerosols and no clouds) and clear sky conditions; thus, the NN models trained on the ClimART dataset would not be suitable for operational applications when the presence of clouds is inevitable.…”

“…With regards to the satisfaction of physical constraints, the previous studies (Krasnopolsky et al, 2010;Lagerquist et al, 2021;Liu et al, 2020;Roh & Song, 2020) trained NN-based emulators to output profiles of heating rates and fluxes at the surface and top-of-atmosphere directly, which causes the issues with energy conservation. Cachay et al (2021) and Ukkonen (2022) chose to predict the radiative fluxes and compute heating rates from fluxes, which ensures physical consistency (Yuval et al, 2021). However, Ukkonen (2022) found that the heating rates are highly sensitive to the continuity in the fluxes profile, and small errors of fluxes lead to relatively large errors in heating rates.…”

Artificial Intelligence for Earth System Predictability (AI4ESP) (2021 Workshop Report)

“…However, it is important to analyze these mature ML frameworks mathematically and test on a set of datasets and computational tasks that represent key challenges in climate science. While efforts such as Cachay et al's (2021) are representative of such ambitions, these benchmarks need to be expanded significantly and swiftly to cover the range of climate science applications.…”

A Physics-Incorporated Deep Learning Framework for Parameterization of Atmospheric Radiative Transfer

“…The previous studies (Krasnopolsky et al, 2010;Lagerquist et al, 2021;Liu et al, 2020;Roh & Song, 2020) trained NN-based emulators to output profiles of heating rates and fluxes at the surface and top-of-atmosphere directly, which causes issues with energy conservation. Cachay et al (2021) and Ukkonen (2022) chose to predict the radiative fluxes and compute heating rates from fluxes, which ensures physical consistency (Yuval et al, 2021). However, Ukkonen (2022) found that the heating rates are highly sensitive to the continuity in the fluxes profile, and minor errors in fluxes lead to relatively large errors in heating rates.…”