Advancing aerosols in Earth system modeling with artificial intelligence

The representations of aerosol and aerosol-cloud interactions (ACI) in conventional Earth system models are overly simplified due to computational constraints. These simple process representations limit the models’ predictive power as they contribute to significant errors in various parts of the simulated climate system. To address this challenge, we developed neural networks to replace aerosol and ACI processes (optics, activation, and precipitation) in the U.S. Department of Energy’s Energy Exascale Earth System Model (E3SM). These neural networks are trained on high-fidelity-high-resolution data and achieve remarkably high accuracy in offline tests. When implemented in E3SM, robust tests and guardrails are needed to ensure that the model produces correct and stable simulations and that their computational cost is low enough so that multi-year global simulations are possible. The hybrid E3SM produces a much more accurate characterization of aerosol and ACI, which leads to a very different climate simulation. To evaluate E3SM, an observationally based emulator has also been developed for understanding aerosol’s role in modulating various atmospheric features across scales in the real world. We highlight that the new hybrid approach, combining physics and artificial intelligence, provides ample opportunities for advancing understanding and predictability of the role of aerosols in the Earth system.