A machine-learning-guided adaptive algorithm to reduce the computational cost of integrating kinetics in global atmospheric chemistry models: application to GEOS-Chem versions 12.0.0 and 12.9.1

Abstract

Abstract. Global modeling of atmospheric chemistry is a great computational challenge because of the cost of integrating the kinetic equations for chemical mechanisms with typically over 100 coupled species. Here we present an adaptive algorithm to ease this computational bottleneck with no significant loss in accuracy and apply it to the GEOS-Chem global 3-D model for tropospheric and stratospheric chemistry (228 species, 724 reactions). Our approach is inspired by unsupervised machine learning clustering techniques and traditional asymptotic analysis ideas. We locally define species in the mechanism as fast or slow on the basis of their total production and loss rates, and we solve the coupled kinetic system only for the fast species assembled in a submechanism of the full mechanism. To avoid computational overhead, we first partition the species from the full mechanism into 13 blocks, using a machine learning approach that analyzes the chemical linkages between species and their correlated presence as fast or slow in the global model domain. Building on these blocks, we then preselect 20 submechanisms, as defined by unique assemblages of the species blocks, and then pick locally and on the fly which submechanism to use in the model based on local chemical conditions. In each submechanism, we isolate slow species and slow reactions from the coupled system of fast species to be solved. Because many species in the full mechanism are important only in source regions, we find that we can reduce the effective size of the mechanism by 70 % globally without sacrificing complexity where/when it is needed. The computational cost of the chemical integration decreases by 50 % with relative biases smaller than 2 % for important species over 8-year simulations. Changes to the full mechanism including the addition of new species can be accommodated by adding these species to the relevant blocks without having to reconstruct the suite of submechanisms.