Learning to Simulate Aerosol Dynamics with Graph Neural Networks.

Area of Science:

• Atmospheric Science
• Computational Chemistry
• Machine Learning

Background:

• Aerosol particles significantly influence climate, weather, and air quality.
• Accurately modeling diverse and evolving particle properties requires computationally intensive particle-resolved models.
• Existing models struggle with the computational cost of capturing detailed aerosol dynamics.

Purpose of the Study:

• To accelerate computationally expensive particle-resolved aerosol models.
• To introduce a novel machine learning framework, Graph-based Learning of Aerosol Dynamics (GLAD), for aerosol simulations.
• To train a surrogate model for predicting aerosol microphysics and chemistry.

Main Methods:

• Implemented a graph-network-based simulator (GNS) where particles are graph nodes.
• Trained a graph neural network (GNN) using output from the PartMC-MOSAIC particle-resolved model.
• Simulated aerosol dynamics, including condensation of sulfuric acid onto mixed-composition particles.

Main Results:

• The trained GNN accurately learned chemical dynamics and generalized across different scenarios.
• Achieved efficient training and prediction times compared to traditional models.
• Demonstrated the robustness and adaptability of the GLAD framework.

Conclusions:

• GLAD offers an efficient and accurate method for simulating aerosol dynamics.
• This machine learning approach can significantly accelerate climate and air quality modeling.
• The framework shows promise for advancing our understanding of aerosol microphysics and chemistry.