For pattern-based simulation methods such as SIMPAT, filtersim, wavesim, ect, patterns are stored by scanning a training image with a sliding template. Dimensional reduction and pattern clustering are then preformed to generate a set of cluster prototypes. During simulation, the conditional data is matched to the best cluster by subsequent similarity measurements and a pattern is randomly sampled from that class. In this paper, a multiple-point, pattern-based, stochastic simulation algorithm using Restricted Boltzmann Machine (RBMSim) is proposed. In the learning phase of RBMSim, the clusters are learned by a contractive divergence learning algorithm which aggregates the two steps of dimensional reduction and pattern clustering. In simulation phase of RBMSim, the two steps of classification and sampling are also aggregated with an efficient Gibbs sampling algorithm. The contributions of this paper are: the learning and simulation are embedded within a mathematically well-defined probabilistic model which represents the entire configuration space; and the learning and simulation phases share the same Markov chain, Monte Carlo based sampling algorithm which explores the entire configuration space efficiently.