An Improved Neural Network for Shortwave Radiative Feedback Quantification

The radiative kernel method has become a key tool in quantifying individual climate feedbacks in the forcing-feedback framework of climate change. This method, however, makes linearity and independence assumptions about climate feedbacks that have been observationally violated in areas where climate change involves feedback variable changes of large magnitudes with complex interactions, such as the Arctic. While neural-network-based methods of feedback quantification, which avoid these assumptions, have been proposed, existing approaches require inputs irrelevant to radiative transfer theory or make physically-inconsistent predictions of radiative sensitivity. In this work, we improve existing neural network methods with data augmentation and Sobolev training. Our new neural network is able to make physically-consistent predictions of top-of-atmosphere net shortwave radiative sensitivity, and achieves better radiative closure than the kernel method in a case study of analyzing shortwave radiative feedbacks in the Arctic.