Abstract： The reactor network synthesis problems are commonly complex non-linear programming (NLP) problems, which are difficult to gain global optimal solutions using the traditional optimization approaches. To avoid the above disadvantage, a two-level optimization algorithm was proposed according to the characters of the reactor network model based on continuous stirred-tank reactors (CSTR). By this algorithm, the NLP of the reactor network synthesis is divided into the linear programming in flow rate and reactor volume space, and the stochastic optimization problem in concentration space. As a result the algorithm reduces the scales and difficulties of the problem. Meanwhile, by applying the global optimization algorithm to the concentration space optimization, the probability of obtaining the global optimal solution is improved. The results of the example studies show that the two-level optimization algorithm proposed can give better structure of the reactor network as well as the reactor type and size in the network.