Exposure to ambient concentrations of ozone (O3) is the second largest pollution-related risk of premature death in the US, leading to approximately 20,000 premature deaths annually. Tens of…

This is a $770,107.72 grant awarded by NASA to the University of Colorado (CU) to develop a computationally efficient chemical data assimilation system for air quality (AQ) modeling and forecasting. CU will build a surrogate model for the chemical solver in the widely used GEOS-Chem AQ model, with the goal of making AQ modeling less computationally intensive. This will enable improved assimilation of satellite observations of atmospheric composition, such as nitrogen dioxide from future geostationary satellites, to enhance AQ forecasting. The work supports NASA's broader Applied Sciences and Earth Science Technology Office objectives for developing data-driven modeling tools. No set-aside designation was used for this research grant award.

Exposure to ambient concentrations of ozone (O3) is the second largest pollution-related risk of premature death in the US, leading to approximately 20,000 premature deaths annually. Tens of millions of US citizens live in areas where O3 concentrations exceed federal standards. While air quality forecasts could minimize these exposure risks and help develop attainment strategies, accurate and reliable O3 forecasts have been elusive owing to the computational complexity of O3 air quality modeling. From the prediction of urban-scale pollution distributions, to short-term O3 forecasts, to better understanding the relationship between O3 and climate change, a persistent challenge in the atmospheric community is the computational expense of chemistry within the models used to research these problems. To address these challenges, this project aims at building a robust and computationally efficient chemical DA system, merging research in compressive sampling and machine learning for large- scale dynamical systems. In particular, we will use low-rank tensor representations, demonstrated recently for the first time by CO-PI doostan for surrogate modeling of chemical kinetics. This approach allows for efficient construction of a surrogate which itself is very computationally cheap to evaluate. Our approach will also draw from expertise in polynomial chaos expansions (PCE a spectral representation of the model solution that can be constructed nonintrusively, i.e., by treating the chemical model as a black box) coupled with multiscale stochastic preconditioners (to address stiffness of the chemical system) to develop fast surrogate models for atmospheric chemistry. In particular, we will accomplish the following objectives: (1) develop, test, and deliver a surrogate model for the chemical solver in a widely used AQ model (geos-chem). (2) generalize the surrogate model generation procedure within a software toolbox applicable to any user-provided chemical mechanisms. (3) demonstrate the benefits of using a surrogate-based AQ modeling framework for assimilation of geostationary observations of atmospheric composition to improve O3 simulations. This project will advance computational tools available for AQ prediction, mitigation, and research. Not only do we aim to deliver a surrogate model for the chemical mechanism of an AQ model used by a very large research community (geos-chem), we will provide a software toolbox for generation of surrogate models for any user-provided mechanism. The potential impacts though are much greater, as improvements in computational expediency could affect research in areas from urban air pollution modeling to longterm studies of chemistry-climate interactions. Eliminating the computational bottleneck associated with chemistry will in turn help to promote the broader use of data assimilation within other AQ forecasting systems. As a case study, we will explore and demonstrate the benefits of surrogate modeling for 4D-VAR assimilation using geostationary measurements of NO2, which in the next few years will be available for the first time over north america, europe, and east asia. This supports a broader goal of facilitating the use of NASA remote sensing data for air quality forecasting here in the US. The proposed work is highly relevant to the aist program and broader goals of the earth science and applied sciences programs. Our work directly responds to the NASA aist proposal solicitation for data-driven modeling tools enabling the forecast of future behavior of the phenomena as well as analytic tools to characterize.