A deep learning approach to fast radiative transfer

Stegmann, Patrick G.; Johnson, Billy E.; Moradi, Isaac; Karpowicz, Bryan M.; McCarty, Will

doi:10.1016/j.jqsrt.2022.108088

Cited by 24 publications

(14 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…An NN emulator that can be used in the RTM was developed some time ago (Chevallier et al., 1998) and was applied to the data assimilation system of the numerical weather prediction (NWP) model (Chevallier et al., 2000). The emulation studies in the RTM are still actively performing (Bue et al., 2019; Liang & Liu, 2020; Stegmann et al., 2022), eventually targeting to the aircraft‐satellite data assimilation in relation to the improvement of forward operator. Recent RTM emulator studies based on clear‐sky simulations have shown a of 1.87–10.88‐fold speedup (Liu et al., 2020) when used with the Rapid Radiative Transfer Model for GCMs (RRTMG; Iacono et al., 2008), and 1.8–3.5‐fold (Ukkonen et al., 2020) and up to 4‐fold (Veerman et al., 2021) for the RRTMG—Parallel scheme (Pincus et al., 2019).…”

Section: Introductionmentioning

confidence: 99%

Benefits of Stochastic Weight Averaging in Developing Neural Network Radiation Scheme for Numerical Weather Prediction

Song

Roh

Lee

et al. 2022

J Adv Model Earth Syst

View full text Add to dashboard Cite

Longwave (LW) and shortwave (SW) radiation physics are important for describing the exchange of energy between the Earth and the Sun. Radiation is a fundamental energy source that determines large-scale atmospheric circulation and consequent physical processes. Accurate calculation involving radiation physics using the line-by-line model (Clough et al., 1992(Clough et al., , 2005 requires high computational burden, rendering it important to develop methods that allow rapid calculation of the radiation process. The recent rapid advances in machine learning techniques have led to the development of neural network (NN) emulators for radiation processes in the two main fields: the radiative transfer model (RTM) and radiation parameterization for the numerical weatherclimate prediction model. An NN emulator that can be used in the RTM was developed some time ago (Chevallier Abstract Stochastic weight averaging (SWA) was applied to improve the radiation emulator based on a sequential neural network (SNN) in a numerical weather prediction model over Korea. While the SWA has advantages in terms of generalization such as the ensemble model, the computational cost is maintained at the same level as that of a single model. In this study, the performances of both emulators were evaluated under ideal and real case frameworks. Various sensitivity experiments using different sampling ratios, activation functions, hidden layers, and batch sizes were also conducted. The emulators showed a 60-fold speedup for the radiation processes and 84%-87% reduction of the total computation. In the ideal simulation, compared to the infrequent radiation scheme by 60 times, SNN improved forecast errors by 5.8%-14.1%, and SWA further increased these improvements by 18.2%-26.9%. In the real case simulation, SNN showed 8.8% and 4.7% improvements for longwave and shortwave (SW) fluxes compared to the infrequent method; however, these improvements decreased significantly after 5 days, resulting in 1.8% larger error for skin temperature. By contrast, SWA showed stable 1-week forecast features with 12.6%, 8.0%, and 4.4% improvements in longwave and SW fluxes, and skin temperature, respectively. Although the use of two hidden layers showed the best performance in this study, it was thought that the optimal number of hidden layers could differ depending on the given problem. Compared to temperature and precipitation observations, all experiments showed a variability of error within 1%, implying that the operational use of the developed emulators is possible.

show abstract

Section: Introductionmentioning

confidence: 99%

Benefits of Stochastic Weight Averaging in Developing Neural Network Radiation Scheme for Numerical Weather Prediction

Song

Roh

Lee

et al. 2022

J Adv Model Earth Syst

View full text Add to dashboard Cite

show abstract

“…These methods are often referred to as hybrid approaches and merge calculations based in physics and statistical methods. For example, low-fidelity physical radiative transfer calculations can be augmented by a neural network to match those of high-fidelity calculations (Brodrick et al, 2021), radiative transfer calculations performed at a subset of wavelengths can be extended across the entire spectral range (Le et al, 2020), or a neural network is used to predict the atmospheric transmittance profile that can then be used in a physical RTM (Stegmann et al, 2022). Using a hybrid approach reduces the dimensionality of the challenge compared to end-toend approaches at the cost of an increased computational burden.…”

Section: Radiative Transfer Speed-up Methodsmentioning

confidence: 99%

Radiative Transfer Speed-Up Combining Optimal Spectral Sampling With a Machine Learning Approach

Mauceri

O’Dell²,

McGarragh³

et al. 2022

Front. Remote Sens.

View full text Add to dashboard Cite

The Orbiting Carbon Observatories-2 and -3 make space-based measurements in the oxygen A-band and the weak and strong carbon dioxide (CO2) bands using the Atmospheric Carbon Observations from Space (ACOS) retrieval. Within ACOS, a Bayesian optimal estimation approach is employed to retrieve the column-averaged CO2 dry air mole fraction from these measurements. This retrieval requires a large number of polarized, multiple-scattering radiative transfer calculations for each iteration. These calculations take up the majority of the processing time for each retrieval and slow down the algorithm to the point that reprocessing data from the mission over multiple years becomes especially time consuming. To accelerate the radiative transfer model and, thereby, ease this bottleneck, we have developed a novel approach that enables modeling of the full spectra for the three OCO-2/3 instrument bands from radiances calculated at a small subset of monochromatic wavelengths. This allows for a reduction of the number of monochromatic calculations by a factor of 10, which can be achieved with radiance errors of less than 0.01% with respect to the existing algorithm and is easily tunable to a desired accuracy-speed trade-off. For the ACOS retrieval, this speeds up the over-retrievals by about a factor of two. The technique may be applicable to similar retrieval algorithms for other greenhouse gas sensors with large data volumes, such as GeoCarb, GOSAT-3, and CO2M.

show abstract

“…In these contexts the radiative transfer integral is approximated using parameterizations. One such optimization parameterizes radiative transfer via linear regression (or, more recently, neural networks, Stegmann et al., 2022) against atmospheric and surface conditions, such as the The Joint Center for Satellite Data Assimilation's Community Radiative Transfer Model (Johnson et al., 2023). In the spectral dimension, radiances can be predicted by sparsely sampling frequencies and integrating via a weighted mean, thereby reducing the number of monochromatic radiative transfer calculations required (Buehler et al., 2010; Moncet et al., 2008).…”

Section: Approximations For Spectral Integrals In Radiative Transfer ...mentioning

confidence: 99%

Sparse, Empirically Optimized Quadrature for Broadband Spectral Integration

Czarnecki,

Polvani,

Pincus

2023

J Adv Model Earth Syst

View full text Add to dashboard Cite

Broadband (spectrally‐integrated) radiation calculations are dominated by the expense of spectral integration, and many applications require fast parameterizations for computing radiative flux. Here we describe a novel approach using a linear weighted sum of monochromatic calculations at a small set of optimally‐chosen frequencies. The empirically‐optimized quadrature method is used to compute atmospheric boundary fluxes, net flux profiles throughout the atmosphere, heating rate profiles, and top‐of‐the‐atmosphere forcing by CO2, in the longwave for clear skies. We evaluate the method against two modern correlated k‐distribution models and find that we can achieve comparable errors with 32 spectral points. We also examine the effect of minimizing different cost functions, and find that in order to accurately represent heating rates and CO2 forcing, these quantities must be included in the cost function.

show abstract

A deep learning approach to fast radiative transfer

Cited by 24 publications

References 18 publications

Benefits of Stochastic Weight Averaging in Developing Neural Network Radiation Scheme for Numerical Weather Prediction

Benefits of Stochastic Weight Averaging in Developing Neural Network Radiation Scheme for Numerical Weather Prediction

Radiative Transfer Speed-Up Combining Optimal Spectral Sampling With a Machine Learning Approach

Sparse, Empirically Optimized Quadrature for Broadband Spectral Integration

Contact Info

Product

Resources

About