2021 Progress Report: Machine-learned atmospheric chemical mechanismsEPA Grant Number: R840012
Title: Machine-learned atmospheric chemical mechanisms
Investigators: Tessum, Christopher , Riemer, Nicole
Institution: University of Illinois Urbana-Champaign
EPA Project Officer: Chung, Serena
Project Period: August 1, 2020 through July 31, 2023
Project Period Covered by this Report: August 1, 2020 through July 31,2021
Project Amount: $399,469
RFA: Chemical Mechanisms to Address New Challenges in Air Quality Modeling (2019) RFA Text | Recipients Lists
Research Category: Air , Air Quality and Air Toxics , Early Career Awards
The objective of the work proposed here is to produce a machine-learned condensed chemical mechanism that provides accurate results in air quality modeling simulations under a wide variety of atmospheric conditions. The proposed work builds on preliminary studies where we have created a neural-network-based chemical mechanism that operates with fewer chemical species and orders-of-magnitude less computational cost than the reference mechanism it was trained on, while reproducing diurnal cycles with low error in 90% of randomly initialized simulations. The proposed project will focus on reducing error in the remaining 10% of cases.
We have coupled the near-explicit gas-phase Master Chemical Mechanism (MCM) with the state-of-art Particle-resolved Monte Carlo Model for Simulation Aerosol Interactions and Chemistry (PartMC-MOSAIC). The coupled model MCM-PartMC/MOSAIC can provide us with a reliable mass-conservative chemical mechanism to train against and give us a platform for the rapid generation of condensed chemical mechanisms as additional chemical detail is included in MCM/PartMC-MOSAIC in future research.
We use Neural ODE to create multi-phase chemistry surrogate model which conserves the mass of all chemical elements. The surrogate model is designed to predict the derivative of concentration which is integrated forward in time by an implicit ODE solver. This Neural ODE framework is expected to stably simulate the autoregressive evolution of main chemical species over extended time periods.
We plan to continue Task 2 to scale up our workflow to train and test the neural network to reduce the error. We will prepare a larger dataset by creating scenarios with Latin hypercube sampled initial conditions and running simulations for these scenarios parallelly. By training on a large and complex dataset, the neural network is expected to have lower error and handle the more difficult and challenging cases.
Journal Articles on this Report : 1 Displayed | Download in RIS Format
|Other project views:||All 1 publications||1 publications in selected types||All 1 journal articles|
||Kelp MM, Jacob DJ, Kutz JN, Marshall JD, Tessum CW. Toward stable, general machine-learned models of the atmospheric chemical system. Journal of Geophysical Research-Atmospheres 2020;125:e2020JD032759.||