A systematic approach was developed for reducing the size as well as the computational cost in the evaluation of diffusion coefficients for mechanisms with mixture-averaged diffusivities by bundling species with similar diffusivities into groups. The systematic reduction was formulated as an integer programming problem and solved efficiently with a greedy algorithm. Reduction error was controlled by a user-specified threshold value, and the algorithm was fully automated. The method was then applied to a 20-species reduced mechanism for ethylene and a 188-species skeletal mechanism for n-heptane. Nine bundled species groups were identified for ethylene, while reduced models with 19, 9, and 3 diffusive species groups were developed for n-heptane in ascending order of reduction errors. Validations of the reduced diffusion models obtained with about 10% reduction error in premixed and nonpremixed flames show good agreement with the detailed model, and the worst case reduction error is close to the user-specified level of 10%. Significant reduction in CPU time was observed in the evaluation of the diffusion terms, while the overall time saving is simulation-dependent due to the existence of other terms, such as the chemical source term, that are not affected by the reduction in the diffusion term. (c) 2006 The Combustion Institute. Published by Elsevier Inc. All rights reserved.