When the nuclear accident occurs, the reliable source terms can provide the data support for emergency response measures effectively. The network toolbox of Matlab can be used to realize the nuclear accident source term inversion based on the BP neural network. To improve the accuracy of nuclear accident source term inversion calculation, several important parameters in nuclear accident source term inversion are studied based on the BP neural network, including the number of hidden layer nodes, the kind of training function, the learning rate, and the number of hidden layers. The results show that the optimal hidden layer node number can be figured out in the single hidden layer, and based on the training time and error, the hidden layer node number selected for further studies is 50. Under the condition of the same parameter settings, the training function ″Trainlm″is more suitable than ″Traingdm″ when the amount of nuclear accident source term data is small. And the inversion calculation accuracy by ″Trainlm″ is higher and the training time is shortened by nearly 35% when the hidden layer node number is 50. The high learning rate and double hidden layers can effectively improve the accuracy of the nuclear accident source term inversion, but the training time relatively increases.