Extensive experimental tests and a computational study of the performance in a cross-flow air classifier have been carried out. A computational fluid dynamics (CFD) package-Fluent-is used to first understand and explain why the cuts or the sharpness of cut of this classifier are not as sharp as they ought to be, and then to optimize the geometry and operational conditions. Flow fields of the classifier under various set-up conditions and geometry were measured by using laser Doppler anemometry (LDA). Using sieve analyses and the HELOS-laser method, the patterns of behaviour of separation parameters such as cut size and sharpness of cut have been investigated at different boundary conditions. Using the Fluent package, a two-dimensional computational fluid dynamics model has been developed. The model is based on the Euler-Lagrangian approach. Different turbulence models have been tested. Both Fluent 4.5, with a structured grid, and Fluent 5.1, with structured and unstructured grids, have been used. Discussions and analyses of the experimental, as well as the computational results, are presented. The simulation with a structured grid shows good agreement with experimental data, except for the sharpness of cut. The reasons of poor performance of the classifier have been found. The geometry is optimized and other conditions were also improved. The performance of the classifier is improved. The experimental observations together with the computed results should increase the depth of understanding of the underlying mechanisms.