Predicting atmospheric optical properties for radiative transfer computations using neural networks

Parameterization of optical properties for liquid cloud droplets containing black carbon based on neural network

This paper introduces a novel back propagation (BP) neural network method to accurately characterize optical properties of liquid cloud droplets, including black carbon. The model establishes relationships between black carbon volume fraction, wavelength, cloud effective radius, and optical properties. Evaluated on a test set, the value of the root mean square error (RMSE) of the asymmetry factor, extinction coefficient, single-scattering albedo, and the first 4 moments of the Legendre expansion of the phase function are less than 0.003, with the maximum mean relative error (MRE) reaching 0.2%, which are all better than the traditional method that only uses polynomials to fit the relationship between the effective radius and optical properties. Notably, the BP neural network significantly compresses the optical property database size by 37,800 times. Radiative transfer simulations indicate that mixing black carbon particles in water clouds reduces the top-of-atmosphere (TOA) reflectance and heats the atmosphere. However, if the volume fraction of black carbon is less than 10-6, the black carbon mixed in the water cloud has a tiny effect on the simulated TOA reflectance.

Neural optimization machine: a neural network approach for optimization and its application in additive manufacturing with physics-guided learning

Neural networks (NNs) are increasingly used in design to construct the objective functions and constraints, which leads to the needs of optimization of NN models with respect to design variables. A Neural Optimization Machine (NOM) is proposed for constrained single/multi-objective optimization by appropriately designing the NN architecture, activation function and loss function. The NN's built-in backpropagation algorithm conducts the optimization and is seamlessly integrated with the additive manufacturing (AM) process-property model. The NOM is tested using several numerical optimization problems. It is shown that the increase in the dimension of design variables does not increase the computational cost significantly. Next, a brief review of the physics-guided machine learning model for fatigue performance prediction of AM components is given. Finally, the NOM is applied to design processing parameters in AM to optimize the mechanical fatigue properties through the physics-guided NN under uncertainties. One novel contribution of the proposed methodology is that the constrained process optimization is integrated with physics/knowledge and the data-driven AM process-property model. Thus, a physics-compatible process design can be achieved. Another significant benefit is that the training and optimization are achieved in a unified NN model, and no separate process optimization is needed. This article is part of the theme issue 'Physics-informed machine learning and its structural integrity applications (Part 1)'.

Usefulness of Automatic Hyperparameter Optimization in Developing Radiation Emulator in a Numerical Weather Prediction Model

To improve the forecasting accuracy of a radiation emulator in a weather prediction model over the Korean peninsula, the learning rate used in neural network training was automatically optimized using the Sherpa. The Sherpa experiment results were compared with two control simulation results using learning rates of 0.0001 and 1 for different batch sizes (full to 500). In the offline evaluation, the Sherpa results showed significant improvements in predicting longwave/shortwave heating rates and fluxes compared to the lowest learning rate results, whereas the improvements compared to the highest learning rate were relatively small because the optimized values by the Sherpa were 0.4756–0.6656. The online evaluation results over one month, which were linked with the weather prediction model, demonstrated the usefulness of Sherpa on a universal performance for the radiation emulator. In particular, at the full batch size, Sherpa contributed to reducing the one-week forecast errors for longwave/shortwave fluxes, skin temperature, and precipitation by 39–125%, 137–159%, and 24–26%, respectively, compared with the two control simulations. Considering the widespread use of parallel learning based on full batch, Sherpa can contribute to producing robust results regardless of batch sizes used in neural network training for developing radiation emulators.

Machine Learning Emulation of Gravity Wave Drag in Numerical Weather Forecasting

We assess the value of machine learning as an accelerator for the parameterization schemes of operational weather forecasting systems, specifically the parameterization of nonorographic gravity wave drag. Emulators of this scheme can be trained to produce stable and accurate results up to seasonal forecasting timescales. Generally, networks that are more complex produce emulators that are more accurate. By training on an increased complexity version of the existing parameterization scheme, we build emulators that produce more accurate forecasts. For medium range forecasting, we have found evidence that our emulators are more accurate than the version of the parametrization scheme that is used for operational predictions. Using the current operational CPU hardware, our emulators have a similar computational cost to the existing scheme, but are heavily limited by data movement. On GPU hardware, our emulators perform 10 times faster than the existing scheme on a CPU.