4:00 PM - 4:15 PM
[AAS02-09] Emulation of Broadband-trained deep learning framework for atmospheric radiative transfer (Longwave)
Keywords:Radiative transfer model, Deep Learning
An atmospheric radiative transfer scheme is one of the most time-consuming components in climate simulation. To accelerate computation, using Deep Learning has been rising in radiation schemes such as feed-forward neural networks (FNN), Convolution neural networks (CNN), and Recurrent neural networks (RNN) predicting flux, heating rate, gas optics, etc (Kransnopolsky et al., 2010; Liu et al., 2020; Roh and Song, 2020; Lagerquist et al., 2021; Yuval et al., 2021; Ukkonen, 2022; Yichen et al., 2023). Preceding studies compared accuracy to various deep learning models and used different data and training parameters in each research. Moreover, training data was usually used for the entire band of the radiative transfer model and different numbers of data.
This study explores the best parametrization condition of the deep learning model that trained each band in MstrnX to speed up the radiative transfer calculation.
MstrnX (Sekiguchi and Nakajima, 2008) is the broadband radiative transfer model developed by the Atmosphere and Ocean Research Institute (AORI). This model has been executed in CReSS(Cloud Resolving Storm Simulator) (Tsuboki and Sakakibara, 2002) and global and regional climate models such as NICAM(Non-hydrostatic Icosahedral Atmospheric Model (Satoh et al., 2008), MIROC (the Model for Interdisciplinary Research on Climate) (Tatebe et al., 2019) etc.
The MstrnX was treated in wavenumber 20 – 5000cm-1 and consisted of 39 bands for Longwave. Absorbing gases were provided for water vapor, ozone, carbon dioxide, methane, nitrous oxide, and carbon monoxide. For training, we used data from the global Copernicus Atmospheric Monitoring Service (CAMS) reanalysis (Inness et al., 2019), the Radiative Forcing Model Intercomparison Project (RFMIP) (Pincus et al., 2016), and short-range forecasts data from The European Centre for Medium-Range Weather Forecasts (ECMWF) (Chevallier et al., 2006). For validation data, we used the Correlated K-Distribution Model Intercomparison Project (CKDMIP) (Hogan and Matricardi, 2020) and AFGL Atmospheric Constituent Profiles (Anderson, 1986). The applied variables for deep learning contain pressure, temperature, water vapor, ozone, carbon dioxide, methane, nitrous oxide, and carbon monoxide for input and flux for output. In this study, we used deep learning model structures including U-net, Deep Neural Networks (DNN), and K-nearest neighbor (KNN).
For band one corresponding to wavenumber region 20-160cm-1, the upwelling flux Mean Squared Error (MSE) result of DNN was the largest in the scenario in CKDMIP and KNN was the smallest, especially in Future, with a value of 0.2. On the other hand, the downwelling flux showed a lower MSE value as against to upwelling flux in all cases, and especially in the Future, DNN showed the smallest value 0.1.
In the presentation, the comparison of each scenario of CKDMIP and region of AFGL using broadband-trained deep learning framework prediction for each deep learning model and to find out the best setting of broadband-trained deep learning will be suggested.
This study explores the best parametrization condition of the deep learning model that trained each band in MstrnX to speed up the radiative transfer calculation.
MstrnX (Sekiguchi and Nakajima, 2008) is the broadband radiative transfer model developed by the Atmosphere and Ocean Research Institute (AORI). This model has been executed in CReSS(Cloud Resolving Storm Simulator) (Tsuboki and Sakakibara, 2002) and global and regional climate models such as NICAM(Non-hydrostatic Icosahedral Atmospheric Model (Satoh et al., 2008), MIROC (the Model for Interdisciplinary Research on Climate) (Tatebe et al., 2019) etc.
The MstrnX was treated in wavenumber 20 – 5000cm-1 and consisted of 39 bands for Longwave. Absorbing gases were provided for water vapor, ozone, carbon dioxide, methane, nitrous oxide, and carbon monoxide. For training, we used data from the global Copernicus Atmospheric Monitoring Service (CAMS) reanalysis (Inness et al., 2019), the Radiative Forcing Model Intercomparison Project (RFMIP) (Pincus et al., 2016), and short-range forecasts data from The European Centre for Medium-Range Weather Forecasts (ECMWF) (Chevallier et al., 2006). For validation data, we used the Correlated K-Distribution Model Intercomparison Project (CKDMIP) (Hogan and Matricardi, 2020) and AFGL Atmospheric Constituent Profiles (Anderson, 1986). The applied variables for deep learning contain pressure, temperature, water vapor, ozone, carbon dioxide, methane, nitrous oxide, and carbon monoxide for input and flux for output. In this study, we used deep learning model structures including U-net, Deep Neural Networks (DNN), and K-nearest neighbor (KNN).
For band one corresponding to wavenumber region 20-160cm-1, the upwelling flux Mean Squared Error (MSE) result of DNN was the largest in the scenario in CKDMIP and KNN was the smallest, especially in Future, with a value of 0.2. On the other hand, the downwelling flux showed a lower MSE value as against to upwelling flux in all cases, and especially in the Future, DNN showed the smallest value 0.1.
In the presentation, the comparison of each scenario of CKDMIP and region of AFGL using broadband-trained deep learning framework prediction for each deep learning model and to find out the best setting of broadband-trained deep learning will be suggested.