Distinguishing Between the Short‐Term Climate Responses to Different Stratospheric Aerosol Injection Latitudes With Explainable Artificial Intelligence

Abstract Stratospheric aerosol injection (SAI), whereby reflective particles are released into the stratosphere to induce cooling, is one possible tool to counteract global warming and its associated risks. However, there is much uncertainty surrounding how SAI would be deployed, as well as the potential of novel and unknown risks and impacts. While most previous research has focused on characterizing the long‐term climate response to SAI, we instead analyze the short‐term response to unilateral injections, within 2 years of deployment, using an explainable artificial intelligence (XAI) framework. Specifically, we train neural networks to predict the latitude of injection in climate simulations with continuous 2‐year cycles of unilateral SAI, using seasonal or annual composites of global 2 m‐temperature or precipitation. The network is able to skillfully determine the latitude of SAI, indicating that there would be distinguishable differences in impacts depending on SAI injection latitude within 2 years of deployment. This distinguishability is generally higher for 2 m‐temperatures than for precipitation. It is strongly modulated by the seasonal cycle of interannual variance, as well as injection size, with the highest distinguishability for injections larger than 8 Tg/year. XAI methods indicate that while distinctive temperature impacts could be identified in both the tropics and midlatitudes, distinguishable precipitation impacts are largely restricted to the tropics.