Of close pertinence to climate engineering is the systematic linear and quadratic nonlinear relationships between the response and the radiative perturbation to the climate system, both being representable by the linear response function (LRF) of the system. By knowing the LRF one would have not only a powerful tool to infer the optimal radiative perturbation to regulate the climate response, but also a means to study the internal dynamical modes of the unperturbed system. We have recently made attempts to estimate the LRF of a coupled climate model, aided by Green’s function experiments wherein every representative location is probed with an isolated forcing. However, estimating the LRF through Green’s function approach is an inverse problem, often suffering the problems of ill-posedness. Thus, searching for a climate engineering solution to reduce the anthropogenic climate change effects presents both a computational and theoretical challenge.

One possible solution to this ill-posed inverse problem is through regularization with a machine learning setup linking climate system perturbations and responses. In this study, we develop a Convolutional Neural Network (CNN) to deal with the imagery system response data and achieve the regularized inversion. We train the CNN with the ensemble data from the Green’s function experiments. Assisted by a densification approach, the developed CNN model can predict the pattern of the forcing that is responsible for the climate response for different scenarios that have not been seen by the CNN during training and validation. This promising result provides a proof-of-concept and a strategy for acquiring the optimal forcing for negating the climate change effects.