Multi-scale modeling system

Understanding stratospheric aerosol injection (SAI) requires connecting physical processes across vastly different scales: from how individual aerosol particles form and grow, to how plumes mix in the stratosphere, to how the resulting changes affect global climate patterns. Traditional modeling efforts treat these scales separately, with no coherent framework linking them. This separation makes it difficult to reduce key uncertainties about SAI's safety and efficacy. Currently, no single modeling framework can resolve these processes at the spatial, temporal, and ensemble scales needed to systematically characterize SAI uncertainty.

The pace and computational cost of conventional climate modeling compound this limitation. Researchers chose the last scenarios for GeoMIP (the Geoengineering Model Intercomparison Project) in 2015, finished the simulations in 2019, and published the first paper in 2021. This multi-year cycle makes it difficult to rapidly explore new injection strategies or systematically map the large uncertainty space associated with SAI. Even once modelers define a scenario, fully resolving the coupled microphysical, dynamical, and radiative processes across scales remains prohibitively expensive at the ensemble sizes required to distinguish robust effects from natural variability.

Rapid progress in artificial intelligence provides an opportunity to rethink this modeling infrastructure. AI-based emulators can learn the behavior of computationally intensive physics-based models and reproduce their dynamics at a fraction of the computational cost. This makes it possible to build a new simulation framework that explores the parameter space of SAI accurately and efficiently while retaining all necessary physical processes.

This activity develops an open-source, integrated AI-driven simulation system that traces how injected aerosols evolve from their formation and growth right after leaving an aircraft to their transport through the stratosphere and their resulting effects on regional climate. The framework links four interacting models that each capture different physical processes: aerosol formation and chemistry at the microphysical scale, turbulence and mixing in plumes, stratospheric circulation and transport, and large-scale atmospheric responses. By connecting these models through machine learning rather than traditional climate modeling, the system can run simulations orders of magnitude faster while maintaining scientific accuracy.

The initial effort focuses on connecting two critical components — aerosol microphysics and atmospheric dynamics — to test whether AI-driven integration can reliably connect particle-scale chemistry to global climate impacts. The pilot's first deliverable is a Python-based model coupling aerosol microphysics to large-scale circulation using an offline method. Success in this phase will validate the approach before the framework expands to include turbulence and plume dynamics.

Once validated, the framework can expand to incorporate additional scales, such as injection-plume turbulence and subgrid mixing, which will be developed in parallel through plume evolution modeling. This will enable researchers to explore how decisions at the injection stage (particle size, altitude, location) propagate through plume evolution and ultimately affect regional precipitation, temperature, and other climate variables, all within timelines relevant to decision-making.