The paper presents some recent advances in the study of the inverse kinetics for subcritical systems. A neural-based approach is adopted to predict the reactivity of the multiplying medium through the analysis of the reactor response to a source pulse. An artificial neural network is designed to infer the subcriticality level through the analysis of power evolution. The training set is computed using an approximate model and its performances are then tested directly on experimental measures, here simulated through a detailed space-energy kinetic model. In order to improve the accuracy of the reactivity estimation, various strategies are proposed and compared, including a multi-transient inversion and the use of different kinetic models for the training. The issue of robustness of the inversion scheme to experimental noise is also addressed.