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

Yao, Yichen; Zhong, Xiaohui; Zheng, Yongjun; Wang, Zhibin

doi:10.1002/essoar.10512741.1

Cited by 4 publications

(5 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Additionally, Ukkonen (2022) showed that bidirectional recurrent NNs (RNNs) are more accurate than the feed-forward NNs (FNNs) for emulating atmospheric radiative transfer. Yao et al (2023) also demonstrated that the bidirectional long short-term memory (Bi-LSTM) achieves the best accuracy in radiative transfer parameterization. Therefore, this study has trained and tested both FC networks and Bi-LSTM models.…”

Section: Ml-based Radiation Emulators and Offline Evaluationsmentioning

confidence: 95%

“…Models A and D, as well as models B and E, have approximately the same number of parameters, respectively. More details about the model training and comparisons can be referenced in Yao et al (2023). All the ML-based emulators have been converted to ONNX and run using the ONNX Runtime library version 1.7.…”

Section: Ml-based Radiation Emulators and Offline Evaluationsmentioning

confidence: 99%

See 1 more Smart Citation

WRF–ML v1.0: a bridge between WRF v4.3 and machine learning parameterizations and its application to atmospheric radiative transfer

Zhong

Yao

et al. 2023

Geosci. Model Dev.

Self Cite

View full text Add to dashboard Cite

Abstract. In numerical weather prediction (NWP) models, physical parameterization schemes are the most computationally expensive components, despite being greatly simplified. In the past few years, an increasing number of studies have demonstrated that machine learning (ML) parameterizations of subgrid physics have the potential to accelerate and even outperform conventional physics-based schemes. However, as the ML models are commonly implemented using the ML libraries written in Python, very few ML-based parameterizations have been successfully integrated with NWP models due to the difficulty of embedding Python functions into Fortran-based NWP models. To address this issue, we developed a coupler to allow the ML-based parameterizations to be coupled with a widely used NWP model, i.e., the Weather Research and Forecasting (WRF) model. Similar to the WRF I/O methodologies, the coupler provides the options to run the ML model inference with exclusive processors or the same processors for WRF calculations. In addition, to demonstrate the effectiveness of the coupler, the ML-based radiation emulators are trained and coupled with the WRF model successfully.

show abstract

Section: Ml-based Radiation Emulators and Offline Evaluationsmentioning

confidence: 95%

Section: Ml-based Radiation Emulators and Offline Evaluationsmentioning

confidence: 99%

WRF–ML v1.0: a bridge between WRF v4.3 and machine learning parameterizations and its application to atmospheric radiative transfer

Zhong

Yao

et al. 2023

Geosci. Model Dev.

Self Cite

View full text Add to dashboard Cite

show abstract

“…The data and source code used for training and testing all the deep learning models in this work is available at https://doi.org/10.5281/zenodo.7213941 (Yao et al, 2022). The source code for the MPAS-A model used in this work is available from https://mpas-dev.github.io.…”

Section: Conflict Of Interestmentioning

confidence: 99%

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

Yao

Zhong

Zheng

et al. 2023

J Adv Model Earth Syst

Self Cite

View full text Add to dashboard Cite

The atmospheric radiative transfer calculations are among the most time‐consuming components of the numerical weather prediction (NWP) models. Deep learning (DL) models have recently been increasingly applied to accelerate radiative transfer modeling. Besides, a physical relationship exists between the output variables, including fluxes and heating rate profiles. Integration of such physical laws in DL models is crucial for the consistency and credibility of the DL‐based parameterizations. Therefore, we propose a physics‐incorporated framework for the radiative transfer DL model, in which the physical relationship between fluxes and heating rates is encoded as a layer of the network so that the energy conservation can be satisfied. It is also found that the prediction accuracy was improved with the physic‐incorporated layer. In addition, we trained and compared various types of DL model architectures, including fully connected (FC) neural networks (NNs), convolutional‐based NNs (CNNs), bidirectional recurrent‐based NNs (RNNs), transformer‐based NNs, and neural operator networks, respectively. The offline evaluation demonstrates that bidirectional RNNs, transformer‐based NNs, and neural operator networks significantly outperform the FC NNs and CNNs due to their capability of global perception. A global perspective of an entire atmospheric column is essential and suitable for radiative transfer modeling as the changes in atmospheric components of one layer/level have both local and global impacts on radiation along the entire vertical column. Furthermore, the bidirectional RNNs achieve the best performance as they can extract information from both upward and downward directions, similar to the radiative transfer processes in the atmosphere.

show abstract

“…The advantages of flexibility (also with regards to vertical grids) can be retained by only replacing the gas optics component with NNs (Ukkonen & Hogan, 2023; Ukkonen et al., 2020;Veerman et al., 2021). Energy conservation, meanwhile, can be ensured by predicting fluxes and computing heating rates from those using a physical equation, which has been combined with a hybrid loss function to minimize heating rate errors (Ukkonen, 2022a; Yao et al., 2023). Although emulators may yet prove useful, for instance as a way of porting code to graphics processing units (GPUs), a recent study (Ukkonen, 2022a) indicates that they suffer from similar speed‐accuracy trade‐offs as traditional radiation schemes: a recurrent NN approach which structurally mimics radiative transfer computations gave much better accuracy than dense networks, but also a smaller speed‐up.…”

Section: Introductionmentioning

confidence: 99%

Twelve Times Faster yet Accurate: A New State‐Of‐The‐Art in Radiation Schemes via Performance and Spectral Optimization

Ukkonen,

Hogan

2024

J Adv Model Earth Syst

View full text Add to dashboard Cite

Radiation schemes are critical components of Earth system models that need to be both efficient and accurate. Despite the use of approximations such as 1D radiative transfer, radiation can account for a large share of the runtime of expensive climate simulations. Here we seek a new state‐of‐the‐art in speed and accuracy by combining code optimization with improved algorithms. To fully benefit from new spectrally reduced gas optics schemes, we restructure code to avoid short vectorized loops where possible by collapsing the spectral and vertical dimensions. Our main focus is the ecRad radiation scheme, where this requires batching of adjacent cloudy layers, trading some simplicity for improved vectorization and instruction‐level parallelism. When combined with common optimization techniques for serial code and porting widely used two‐stream kernels fully to single precision, we find that ecRad with the TripleClouds solver becomes 12 times faster than the operational radiation scheme in ECMWF's Integrated Forecast System (IFS) cycle 47r3, which uses a less accurate gas optics model (RRMTG) and a more noisy solver (McICA). After applying the spectral reduction and extensive optimizations to the more sophisticated SPARTACUS solver, we find that it’s 2.5 times faster than IFS cy47r3 radiation, making cloud 3D radiative effects affordable to compute in large‐scale models. The code optimization itself gave a threefold speedup for both solvers. While SPARTACUS is still under development, preliminary experiments show slightly improved medium‐range forecasts of 2‐m temperature in the tropics, and in year‐long coupled atmosphere‐ocean simulations the 3D effects warm the surface substantially.

show abstract

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

Cited by 4 publications

References 24 publications

WRF–ML v1.0: a bridge between WRF v4.3 and machine learning parameterizations and its application to atmospheric radiative transfer

WRF–ML v1.0: a bridge between WRF v4.3 and machine learning parameterizations and its application to atmospheric radiative transfer

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

Twelve Times Faster yet Accurate: A New State‐Of‐The‐Art in Radiation Schemes via Performance and Spectral Optimization

Contact Info

Product

Resources

About