Abstract View
Deep Learning for Prediction of Multiphase Isoprene Oxidation Products
MUNKHZAYA BOLDBAATAR, Mohamadamin Tavakoli, Dora Kadish, Karla Rojas Garcia, Pierre Baldi, David van Vranken, Sergey Nizkorodov, Annmarie Carlton, University of California, Irvine
Abstract Number: 66
Working Group: Aerosol Chemistry
Abstract
Isoprene is the most abundantly emitted biogenic volatile organic compound in the atmosphere and its oxidation with atmospheric oxidants contributes substantially to the global secondary organic aerosol budget. Isoprene oxidation has been studied for many years, but despite the extensive studies, our mechanistic understanding is far from complete and the precise chemical identities of high molecular weight compounds of isoprene oxidation remain speculative. To understand the full extent of atmospheric reactions and the identity of isoprene oxidation products, we approach this problem from the nexus of computer science and chemistry by training and applying an intelligent system “Reaction Predictor.” Reaction Predictor, or shortly RP, is a machine learning-based system that utilizes a deep learning algorithm to predict the outcome of chemical reactions. Reactions are predicted at the level of elementary mechanistic steps that can be chained together to yield complex global reactions. RP is currently trained with 677 radical reactions (resonance, homolysis, abstraction, addition, recombination, hydride shift) and more than 12,000 polar reactions and reactions of atmospherically relevant gas-phase VOCs (α-pinene and isoprene). We apply RP to isoprene oxidation by a hydroxyl radical, OH. Here, we present our RP predictions with mass spectra from laboratory smog chamber experiments, with a focus on chemically identifying large molecular weight products.