A natural object such as a flame with smoke is featured in terms of ambiguous boundaries and complex motion. One way of modeling natural objects is by particle systems, which need a large amount of computation time to calculate interactions among the particles. This paper describes another way of modeling particle motions based on a cellular automaton. Cellular automata are massively parallel computation models that can simulate complex phenomena. In our model, particle motions are simulated in a cellular space with a Margolus neighborhood, which has good conservation properties and collision detectability. This paper shows several applications in two-dimensional cellular space.