An adaptive auto-reduction solver for speeding up integration of chemical kinetics in atmospheric chemistry models: implementation and evaluation within the Kinetic Pre-Processor (KPP) version 3.0.0

Kinetic integration of chemical mechanisms is a central process and major computational bottleneck in the modeling of atmospheric chemistry. Chemical mechanisms for models typically include hundreds of coupled species and are stiff systems requiring the implicit solution of the coupled system of kinetic differential equations. The Kinetic Pre-Processor (KPP) software is a versatile software tool that takes as input a list of species, reactions, and rate coefficients and generates computationally efficient Fortran 90, C, or MATLAB code for solving the stiff system. We present a new version 3.0.0 of the KPP software with a range of improvements over previous versions in performance, diagnostics, versatility, and community openness. We also implement in KPP 3.0.0 a new adaptive auto-reduction solver which decreases the size of any chemical mechanism locally and on the fly where the full complexity of the mechanism is not needed, by partitioning species as "fast" or "slow" based on local production and loss rates. The adaptive solver is implemented with little overhead by cropping the Jacobian matrix locally to remove the "slow" species with no change in memory allocation, and allowing for reuse of the pre-computed matrix terms of the full mechanism. We apply this adaptive solver to the GEOS-Chem global 3-D model of atmospheric chemistry through KPP 3.0.0 and demonstrate a 32% reduction in solver time while maintaining a mean error lower than 1% for key species in the troposphere.