Quantification of Shortwave Surface Albedo Feedback Using a Neural Network Approach

Abstract

Radiative transfer is a nonlinear process. Despite this, most current methods to evaluate radiative feedback, such as the kernel method, rely on linear assumptions. Neural network (NN) models can emulate nonlinear radiative transfer due to their structure and activation functions. This study aims to test whether NNs can be used to evaluate shortwave radiative feedbacks and to assess their performance. This study focuses on the shortwave radiative feedback driven by surface albedo. An NN model is first trained using idealized cases, simulating truth values from a radiative transfer model via the partial radiative perturbation method. Two heuristic cases are analyzed: univariate feedback, perturbing the albedo; and bivariate feedback, perturbing the albedo and cloud cover concurrently. These test the NN’s ability to capture nonlinearity in the albedo–flux and albedo–cloud–flux relationships. We identify the minimal NN structure and predictor variables for accurate predictions. Then, an NN model is trained with realistic radiation flux and atmospheric variable data and is tested with respect to its predictions at different order levels: zero-order for the flux itself, first-order for radiative sensitivity (kernels), and second-order for kernel differences. This paper documents the test results and explains the NN’s ability to reproduce the complex nonlinear relationship between radiation flux and different atmospheric variables, such as surface albedo, cloud optical depth, and their coupling effects.