https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025JD044319
*Authors*: Heng Quan, Daniel D. B. Koll, Nicholas Lutsko, Janni Yuval First published: *04 December 2025* https://doi.org/10.1029/2025JD044319 *Abstract* Solar geoengineering via stratospheric aerosol injection (SAI) poses an optimization problem. How exactly should aerosol be deployed to maximize its benefits while minimizing undesirable side-effects, such as shifts in rainfall patterns? Previous work explored this problem using feedback control based on linear algorithms. Here, we investigate an alternative approach, which also naturally incorporates feedback. We let a reinforcement learning (RL) algorithm control the distribution of stratospheric aerosol concentration in an idealized global climate model (GCM). Within several dozen GCM simulations, RL learns to produce stable and plausible strategies. RL also learns that the optimal geoengineering strategy depends on the time when geoengineering is initiated, which we further explain using a simple energy-balance model. Our results provide a first proof-of-concept that RL can identify promising SAI strategies. *Plain Language Summary* Society might be able to temporarily mitigate the worst impacts of climate change by injecting reflective aerosols into the stratosphere. What is the best way to deploy aerosol without creating additional problems, such as disrupting monsoons and storm tracks? Previous research tackled this question using linear algorithms from the control literature. Our goal is to investigate the potential of an alternative algorithm class, namely an AI technique called reinforcement learning (RL). We train an RL algorithm to control the pattern of stratospheric aerosol concentration inside a global climate model. Initially, the algorithm produces random aerosol patterns. Over time, it learns how to best use aerosol to keep temperature and rainfall patterns in the model close to a desired target state. The algorithm also learns nonobvious strategies, such as how to vary the concentration of aerosol over time to better overcome Earth's thermal inertia and cool the climate faster. These results show RL is a feasible and promising technique for future geoengineering research. More work is needed to compare the strategies identified here to those produced by alternative algorithms. *Key Points* We let a reinforcement learning (RL) algorithm control the stratospheric aerosol in a global climate model (GCM) RL learns to produce stable and plausible stratospheric aerosol injection strategies within several dozen GCM simulations RL shows that the optimal geoengineering strategy depends on the time when geoengineering is initiated *Source: AGU* -- You received this message because you are subscribed to the Google Groups "geoengineering" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/geoengineering/CAHJsh995gvb36MMDbWXxVD3_cutdxThqSXbiFyJLrM2CGL%2BsGA%40mail.gmail.com.
