A neural network method has been developed to analyze the climate feedbacks. The performance of the method was evaluated by comparison of radiation flux against the ERAi reanalysis data and comparison of the feedbacks against the kernel method. We find that the NN method can well reproduce the radiation fluxes and their interannual anomalies, in terms of both global mean and distribution. The global mean feedbacks analyzed from the two methods are generally in agreement. However, the NN method possesses a few notable advantages:
(i) The NN method accounts for the nonlinear effect in the feedbacks. Such nonlinear effect corresponds to the second and higher order terms of the Taylor expansion of ∆R, which are neglected in the kernel method. This effect is shown to suppress the radiation response to large perturbations in the surface albedo caused by Arctic sea ice melt (see figure). The NN method, without making linear extrapolation in feedback quantification, generally measures a more moderate albedo feedback than the kernel method.
(ii) The NN method can directly measure the cloud feedback and its components. We find that the NN method reproduces the distribution of the cloud feedback from the kernel method.
Area of Research: Climate
About the Author / Biography
Tingting Zhu is a Ph.D student under Prof. Haikun Wei at School of Automation, in Southeast University China, and as a visiting student under Prof. Yi Huang at department of Atmospheric and Oceanic Sciences, in McGill University. During the visiting period, she was awarded by the Fonds de recherche Nature et technologies of Quebec and the China Scholarship Council. Her current research interests include solar power generation forecast, machine learning and climate feedback.