In this paper, a hierarchical methodology for the design of manufacturing cells is proposed, synthesizing the capabilities of new pattern recognition methods for rapid clustering of large part-machine data sets, with multi-objective optimization capabilities of mathematical programming. The procedure includes three phases. In Phase I, part families and associated machine types are identified through neural network methods for pattern recognition. Phase II is a cell formation phase that involves the assignment of part families and individual machines to create independent cells. It takes into account several factors such as capacity constraints, cell size restrictions, minimization of load imbalances and provision of flexibility, Phase III attempts to minimize inter-cell traffic further for families that may still have to be processed in more than one cell. The methodology is illustrated using several examples.