The separation property of Liquid State Machine (LSM) is a key for its power of computing, but the weights and delays of the inter-connections in the spiking neural circuit are usually randomly created and kept unchanged, which hinders the performance of the LSM greatly. In this paper, particle swarm optimization (PSO) was applied to optimize the weights and delays of the circuit so as to enhance the separation property of the LSM. Separation of random spike trains and Fisheriris data-set classification experiments are done by the optimized circuit. Demonstration examples show that the PSO can enlarge the separation property of the circuit greatly compared to the normal Hebbian-learning algorithm and enhance the computing ability of LSM.