Skip to main content
U.S. flag

An official website of the United States government

A Neural Network Aerosol Optics Emulator for E3SM

Presentation Date
Tuesday, January 30, 2024 at 2:30pm - Tuesday, January 30, 2024 at 2:45pm
The Baltimore Convention Center - 345/346



Climate models must accurately represent the direct radiative effects of simulated aerosol populations, but directly computing aerosol optical properties is computationally expensive. This work develops and tests two neural network emulators capable of estimating the bulk optical properties of populations of both homogeneous spherical particles and coated spheres, that can be used to replace the existing parameterization in the Energy Exascale Earth System Model’s (E3SM) Atmosphere Model (EAM). The homogeneous sphere model reproduces simulated bulk optical properties from Mie scattering code with only negligible error and provides a significant improvement in accuracy over the existing parameterization. Meanwhile, the core-shell model introduces the capacity to represent sulfate coated black carbon and moves beyond the capabilities of the existing parameterization. The neural networks leverage symmetry in the Mie calculations to provide accurate solutions for a large range of particle size and wavelength combinations and are well suited for integration into other models. We will discuss the neural network development process and share initial results from integration with E3SM.

23rd Conference on Artificial Intelligence for Environmental Science
Funding Program Area(s)
Additional Resources:
NERSC (National Energy Research Scientific Computing Center)