CHEMULATION - Emulating atmospheric chemistry with machine learning

AI Technology: Machine Learning, Deep Learning.

Atmospheric chemistry represents a major part of the computational burden of current atmospheric models, notably due to the stiffness of the corresponding ODE system that forces the use of computationally expensive implicit numerical solvers.

The research group at BSC is engaged in the development of AI-based emulators of the atmospheric chemistry.

When integrated in weather-chemistry models like MONARCH, they are expected to considerably sped-up the calculations, enhancing the capabilities for refining our representation of the chemistry in current models (especially relevant in chemistry-climate models where it remains very simplified). 

Using BSC's in-house CAMP multiphase library, we already prepare large training datasets (1 million chemical simulations) for a few chemical mechanisms. 

The resources on EuroHPC systems will allow initiating the development of a first baseline fully coneccted deep learning model, to explore the impact of the dataset size on performance, to investigate the introduction of physical constaints and finally to explore more sophisticated architectures including graph neural networks and neural operators.