Data-driven emulation of modal aerosol microphysics via neural operator-based modeling.

Area of Science:

• Earth System Science
• Atmospheric Science
• Computational Science

Background:

• Aerosol microphysical processes are complex and operate at small scales, posing challenges for accurate Earth system simulations.
• Existing models struggle with the computational demands of simulating these processes at regional and global scales.

Purpose of the Study:

• To develop and evaluate a surrogate model, the aerosol deep operator network (ADON), for emulating aerosol microphysics parameterization.
• To improve the accuracy and efficiency of Earth system models like the Energy Earth System Model version 2 (E3SMv2).

Main Methods:

• Constructed a physics-inspired dual-net architecture for the ADON surrogate model.
• Trained the model on a large dataset (9.8 million samples) from E3SMv2 simulations under cloud-free conditions.
• Incorporated spatial, temporal, and principal component features into the dual-net architecture.

Main Results:

• The ADON model achieved high accuracy, with R-squared scores over [Formula: see text] for lognormal aerosol modes.
• The model effectively captured aerosol representations and their relationships with atmospheric variables.
• Analysis revealed feature importance, highlighting key input variables impacting predictive capacity.

Conclusions:

• The validated ADON model demonstrates significant efficiency for online inference on CPUs and GPUs.
• ADON shows strong potential for robust predictive modeling in large-scale Earth system computations.
• This surrogate model offers a pathway to more accurate and efficient climate simulations.