RadNet 1.0: exploring deep learning architectures for longwave radiative transfer

Abstract

Abstract. Simulating global and regional climate at high resolution is essential to study the effects of climate change and capture extreme events affecting human populations. To achieve this goal, the scalability of climate models and efficiency of individual model components are both important. Radiative transfer is among the most computationally expensive components in a typical climate model. Here we attempt to model this component using a neural network. We aim to study the feasibility of replacing an explicit, physics-based computation of longwave radiative transfer by a neural network emulator and assessing the resultant performance gains. We compare multiple neural-network architectures, including a convolutional neural network, and our results suggest that the performance loss from the use of conventional convolutional networks is not offset by gains in accuracy. We train the networks with and without noise added to the input profiles and find that adding noise improves the ability of the networks to generalise beyond the training set. Prediction of radiative heating rates using our neural network models achieve up to 370× speedup on a GTX 1080 GPU setup and 11× speedup on a Xeon CPU setup compared to the a state-of-the-art radiative transfer library running on the same Xeon CPU. Furthermore, our neural network models yield less than 0.1 K d−1 mean squared error across all pressure levels. Upon introducing this component into a single-column model, we find that the time evolution of the temperature and humidity profiles is physically reasonable, though the model is conservative in its prediction of heating rates in regions where the optical depth changes quickly. Differences exist in the equilibrium climate simulated when using the neural network, which are attributed to small systematic errors that accumulate over time. Thus, we find that the accuracy of the neural network in the “offline” mode does not reflect its performance when coupled with other components.

Improved Weather Forecasting Using Neural Network Emulation for Radiation Parameterization

“…The radiative transfer for TOVS (RTTOV) has been developed using multiple linear regression since 1999, and it has been widely used in data assimilation in the NWP model (Saunders et al, 2018). Recent studies on radiative transfer modeling have extended the application of various AI techniques, including multiple linear regression, deep neural network (DNN), adaptive network-based fuzzy inference system, and convolutional neural network (CNN) for radiation processes over a clear sky (Bilgic & Mert, 2021;Liu et al, 2020;Ukkonen et al, 2020;Veerman et al, 2021) and 3-dimensional cloud radiative effects (Meyer et al, 2021). As these studies do not utilize repetition by time integration, such as in the numerical forecast model, errors caused by emulation do not accumulate.…”

“…Pal et al (2019) reported that a DNN radiation emulator can produce an 8-to 10-fold speedup and 90%-95% accuracy in the Super-Parameterized Energy Exascale Earth System Model from the United States Department of Energy; however, they did not provide a specific reduction in the total computational cost. Liu et al (2020) showed that the use of the CNN emulator could reduce the RMSE for clear-sky LW cooling rates by 41%-59%, compared to a DNN emulator with three hidden layers; however, the CNN resulted in an approximately 10-fold slowdown in contrast to a 10.88-fold speedup for the DNN (i.e., the CNN was ∼100 times slower than the DNN). We can see that the use of deep hidden layers or more complex structures (Liu et al, 2020;Pal et al, 2019) may not always produce better performance compared with an NN with a single hidden layer (in terms of speedup), although it offers a variety of possibilities for optimization.…”

“…Liu et al (2020) showed that the use of the CNN emulator could reduce the RMSE for clear-sky LW cooling rates by 41%-59%, compared to a DNN emulator with three hidden layers; however, the CNN resulted in an approximately 10-fold slowdown in contrast to a 10.88-fold speedup for the DNN (i.e., the CNN was ∼100 times slower than the DNN). We can see that the use of deep hidden layers or more complex structures (Liu et al, 2020;Pal et al, 2019) may not always produce better performance compared with an NN with a single hidden layer (in terms of speedup), although it offers a variety of possibilities for optimization. Belochitski and Krasnopolsky (2021) also noted that the risks of developing the DNN emulator associated with the control of complexity and nonlinearity.…”

Compound Parameterization to Improve the Accuracy of Radiation Emulator in a Numerical Weather Prediction Model

“…Recent advances in machine‐learning techniques have provided new opportunities to significantly accelerate the computation speed of numerical weather prediction (NWP) models. Among the various fields of numerical models, radiation physics for longwave (LW) and shortwave (SW) accounts for the most significant computational burden in models and has the oldest history of developing machine‐learning emulators in radiative transfer modeling (Chevallier et al., 1998; Liu et al., 2020; Meyer et al., 2021; Ukkonen et al., 2020; Veerman et al., 2021), data assimilation (Chevallier et al., 2000), and radiation parameterization in climate simulation models (Belochitski et al., 2011; Krasnopolsky et al., 2005, 2010; Pal et al., 2019) and NWP models (Roh & Song, 2020; Song & Roh, 2021). In previous studies, acceleration of two orders was achieved by replacing the radiation parameterization with a neural network (NN) emulator for numerical models.…”

Applying Deep Learning to Clear-Sky Radiance Simulation for VIIRS with Community Radiative Transfer Model—Part 2: Model Architecture and Assessment

“…Using simple, statistical, nonlinear approximation instead of a complicated physical-based model in ANNs renders a more computationally efficient method to achieve a similar job to that of the physical-based model without significant accuracy loss [13][14][15][16][17][18]. These advantages have attracted an increasing number of remote sensing scientists to explore the possible replacement of the radiative transfer (RT) forward model or inversion with the ANN model in recent years [19][20][21][22][23][24][25][26]. Given the complicated nature of the RT model and its input, emulating a full RT model using only one ANN architecture is currently impossible.…”