Source term inversion of short-lived nuclides in complex nuclear accidents based on machine learning using off-site gamma dose rate

During nuclear accidents, large amounts of short-lived radionuclides are released into the environment, causing acute health hazards to local populations. Therefore, it is particularly important to obtain source-term information to assist nuclear emergency decision makers in determining emergency protective measures. However, it is extremely difficult to obtain reliable contaminant monitoring instrument readings to estimate the source term based on core conditions, release routes, and release conditions. Currently, a wide variety of source-term inversion methods are attracting increasing attention. In this study, the release rates of four typical short-lived nuclides (Kr-88, Sr-91, Te-132, I-131) in two complex nuclear accident scenarios were estimated using a machine-learning method. The results show that the best estimation performance is obtained with the long short-term memory network, and the mean absolute percentage errors for the release rates of the four nuclides at 10 h under the two nuclear accidents are 9.87% and 11.08%, 17.49% and 16.51%, 7.16% and 8.35%, and 38.83% and 41.87%, respectively. Meanwhile, the mean absolute percentage errors for Te-132 (7.16% and 8.35%) were the lowest among all the estimated nuclides. In addition, stability analysis showed that the gamma dose rate was the key parameter affecting the estimation accuracy.