This work presents the feasibility studies of mustard gas recognition using neural network algorithms and data samples measured by a novel, non-invasive device, which is being developed at the Jagiellonian University within the framework of the SABAT (Stoichiometry Analysis By Activation Techniques) project. The data samples used to train and validate the performance of the used algorithms were based on realistic Monte Carlo simulations which formed histograms of energy depositions for three intervals of detection gamma quanta time. Multiple neural network models have been trained and tested using 7-folds cross-validation in order to analyze how does detector?s sensitivity impact model?s precision. The best results have been achieved for the LaBr3:Ce detector. Nonetheless, based on simulated data from inexpensive, less sensitive NaI:Tl detector obtained model?s accuracy has been only slightly lower. It is expected, that training a model on a larger dataset, without a burden of correlation caused by a limitation in the size of the simulations, may improve the results for NaI:Tl detector.