American Association for Aerosol Research - Abstract Submission

AAAR 38th Annual Conference
October 5 - October 9, 2020

Virtual Conference

Abstract View


Application of Machine Learning to Development of Atmospheric Chemical Mechanisms

YUANLONG HUANG, Weimeng Kong, John Seinfeld, California Institute of Technology

     Abstract Number: 568
     Working Group: Instrumentation and Methods

Abstract
Atmospheric chemistry plays an important role in climate change and human health. The temporal evolution of species emitted to or formed in the atmosphere is evaluated by three-dimensional chemical transport models (CTMs) simulating simultaneous chemical reactions and transport. The number of all the species comprising atmospheric chemistry can be immense. Moreover, representing the detailed steps that proceed from initial oxidation of a volatile organic compound (VOC) with the principal oxidants, OH, O3, or NO3, to important products is a major challenge in atmospheric chemistry. Generally, chemical mechanisms employed in atmospheric models are limited to a few hundred species and reactions, which still can require up to 90% of the overall computational resources for solving the overall chemical transport model. This is a severe limitation on the ability to simulate atmospheric chemistry. Machine learning has been demonstrated to be successful in handling complex simulations by significantly reducing computational burdens but keeping precision. Here we explore the potential of employing machine learning in atmospheric chemical mechanisms to reproduce with fidelity the molecular routes leading from compounds emitted into the atmosphere to ultimate oxidation products, including aerosol species. Chemical reaction networks generated by machine learning offer the opportunity not only to replace existing mechanisms used in atmospheric models but also to be computationally efficient. Machine learning-derived reaction mechanisms based on the explicit atmospheric chemical mechanisms (e.g., isoprene oxidation) that have been developed in the laboratory and theoretical studies will be trained by state-of-the-art long short-term memory networks.