Designing Optimal Forcings for SRM Using Differentiable AI Climate Models

Determining where and when to reflect sunlight to reduce extreme heat, while also minimizing harmful side effects, is a major challenge with the study of sunlight reflection methods (SRM).

This challenge is heightened by the fact that traditional climate models cannot easily calculate how subtle changes in an SRM strategy would affect climate outcomes, hindering systematic exploration of effective SRM design.

In this project, researchers will use novel AI-based climate models to overcome this limitation, minimizing the harmful impacts of an SRM deployment on rainfall patterns. These models are built using a neural network architecture and trained with methods that allow researchers to better calculate how small changes in inputs affect outputs—leading to a more robust understanding of how a small change in an SRM strategy might lead to changes within a climate system.

The study will focus on two SRM approaches: stratospheric aerosol injection (SAI), which entails dispersing tiny reflecting particles of sulfur dioxide into the stratosphere; and marine cloud brightening (MCB), which seeks to increase the reflectivity of clouds over the ocean.

The climate models being used for this project are NeuralGCM and LUCIE-3D, which represent the two primary approaches to AI climate modeling: one that is informed by physical laws directly incorporated into the model, and one that is primarily data driven.