AAAR 36th Annual Conference October 16 - October 20, 2017 Raleigh Convention Center Raleigh, North Carolina, USA
Abstract View
Development of a CMAQ Adjoint Model with Aerosol Capabilities
SHUNLIU ZHAO, Amir Hakami, Matt Turner, Daven Henze, Shannon Capps, Peter Percell, Jaroslav Resler, Jesse Bash, Sergey Napelenok, Kathleen Fahey, Rob Pinder, Armistead G. Russell, Athanasios Nenes, Jaemeen Baek, Gregory Carmichael, Charles Stanier, Adrian Sandu, Tianfeng Chai, Daewon Byun, Carleton University
Abstract Number: 487 Working Group: Regional and Global Air Quality and Climate Modeling
Abstract An adjoint air quality model provides location- and time-specific gradients of an air quality metric to model inputs and lends itself to backward sensitivity analysis, source apportionment, optimal pollution control, data assimilation and inverse modeling for scientific and policy applications. A gas-phase adjoint model for CMAQ was previously developed (Hakami et. al, 2007) and has been used in various applications related to ozone. However, the lack of aerosol and cloud processes in the adjoint model has so far prevented applications related to aerosols, which in turn has imposed significant limitation on multi-pollutant applications on topics such as human health and climate. A collaborative effort has been underway for the past few years to develop a full adjoint version for CMAQ. In this work, we will present the development work and provide example applications.
The adjoint model development has been assisted with different Automatic Differentiation (AD) tools for different processes. Code pre-processing is required for AD and one example is to address the problem associated with the bisection procedure in the aerosol thermodynamics module, ISORROPIA, and the secondary organic aerosol module. The adjoint code generated by AD was originally evaluated on a process-by-process basis against the Finite Difference Method (FDM). The FDM, which has long been used for verification of forward (DDM) or backward sensitivity modules, repeatedly failed to produce reliable sensitivity estimations. We have instead moved to using the Complex Variable Method (CVM) for evaluation of the adjoint code where the FDM proves problematic. A continuous adjoint of advection was developed manually which is superior in certain applications. Finally, we will discuss various applications that the CMAQ-adjoint model can be used for, provide examples in using the adjoint model to address policy and health questions, and discuss the specific pre-processing modules that have been developed for such applications.