A coal gangue pile in Fushun City was selected as the study area, and the groundwater numerical simulation model for this area was established based on physical condition. After that, we applied the model in predicting possible spatial and temporal variation of groundwater pollutant in research area. Based on forward modelling, a hypothetical example was designed to explore the application of simulation-optimization method and simultaneously carried out inverse identification of groundwater pollution sources and model parameters. The Kriging method and BP Neural Network method were proposed to establish a surrogate model of the groundwater numerical model so as to reduce the computational load caused by the repeated invocation of the simulation model by the optimization model, and the surrogate model was then integrated to simulation-optimization model where simulated annealing method was used. The study reached the following conclusions:As the surrogate model of the simulation model established by Kriging method, the mean relative error of the output concentration is 0.3%, by contrast with 1.5% with BP Neural Network method. The identification error of the release intensity of pollution source with two methods are both below 0.5%, and the identification error of hydraulic conductivity both partitions is no larger than 5%. Above all, the accuracy of surrogate model established by Kriging method is higher than BP neural network method. It is proved that the inverse identification by using the simulation-optimization method based on two surrogate models is effective and accurate.