Abstract: Inverse identification of groundwater contamination source involves the reconstruction of pollution source features using limited and discrete observation data. However,when the simulation-optimization method is used to perform the identification task,the numerical simulation model must be invoked repeatedly,inevitably leading a large computational load. Moreover,when optimization model is solved,traditional particle swarm optimization algorithm commonly falls into the local minimum,which hinders precise identification. In the present study,a hypothesis case is designed to evaluate the performance of proposed framework. The numeric simulation model is embedded as equation constraint of optimization model where the objective function is the bias between simulation outputs and observations and the decision variables denotes features of contamination source. In particular,the optimization model is established to simultaneously estimate the release history of three potential contamination sources and hydraulic conductivity. To reduce the huge calculated burden,the BP neural network is introduced to substitute the simulation model. Furthermore,to alleviate being trapped into local minimum,adaptive weighted strategy is proposed to improve the particle swarm algorithm. The identification results indicate that:（1） The BP neural network surrogate model can approximate the input-output relationship of the simulation model with desired accuracy of R square of 0.99,and the running speed of surrogate is evidently swifter than that of the numerical simulation model.（2） Compared with the traditional particle swarm optimization algorithm,the adaptive weighted particle swarm optimization algorithm can substantially improve the convergence speed and optimal efficiency by adjusting the parameters and iteration termination conditions of the optimization algorithm,and the relative error of the optimal solution is less than 5%.