To estimate the severity of nuclear accident, a back propagation (BP) neural network basic model is built for source term inversion during a nuclear accident. The release rates of I-131, Cs-137, Xe-133 and Kr-85 are selected as target signals, and the Matlab software is used to perform the calculations for source term inversion. The results show that in a single hidden layer, the train mean square error decreases firstly but increases thereafter with increasing the number of nodes from 5 to 60, and reaches the minimum value of 41.1% when the number of nodes is 25. Increasing the learning rate from 0.01 to 0.2 can reduce the relative error variance for each nuclide. The relative errors of release rates of I-131, Cs-137, Xe-133 and Kr-85 are 6.7%, 49.7%, 92.3% and 92.0%, respectively, when the learning rate is 0.2. The source term inversion is tested at the node number of 25 and the learning rate of 0.2, and the results show that the relative test errors of release rates of I-131, Cs-137, Xe-133 and Kr-85 are 56.7%, 49.1%, 92.4% and 92.0%, respectively.