Layered composite beams exhibit significant spatial variabilities in their properties due to the variations in the individual lamina properties and the constituent matrix.The focus of this study is on the development of a stochastic finite element framework using methods of stochastic multi-scale analysis for quantification of the uncertainties in composite structures. The variations in the individual laminae stiffness properties are modeled as random fields. The finite element (FE) method involves two stages of discretization. The displacement fields are discretized into variables along the spatial dimension using traditional FEM. The random fields are discretized into a vector of correlated random variables along a mathematical dimension using polynomial chaos expansions. As the random field discretization is carried out for each individual laminae, the number of random variables associated with the problem becomes large, making uncertainty quantification computationally prohibitive. Reduction in the number of random variables is achieved by subsequently considering the problem in the structure scale rather than the lamina scale. Equivalence in the uncertainty between the two scale models is ensured by solving an optimization problem. This is achieved by first estimating the probabilistic statistics of the structure response of the fine scale model. Next, the stochastic response is represented in terms of a structure scale model using PCE, whose dimension and basis functions are determined such that the pdf of the response obtained from both models are equivalent. This equivalence is achieved using genetic algorithms. The coarse scale model can be subsequently used for reliability and life estimation studies. The developments are illustrated through a numerical example involving a composite beam. © 2013 Taylor & Francis Group, London.