Unfolding of fast neutron spectra by superheated drop detectors using Adaptive Network-Based Fuzzy Inference System (ANFIS)

Abstract

Abstract ANFIS (Adaptive Network-Based Fuzzy Inference System) was utilized for fast neutron spectrometry using responses of a set of superheated drop detectors under different external pressures (different response functions). The designed ANFIS with the best performance was trained and tested by a set of neutron spectra and corresponding detector readings. The neutron spectra were the target data for the ANFIS and the corresponding responses of five superheated drop detectors were the input data. 90% of input and corresponding target data were used for training and the rest for testing. Hybrid algorithm was used for the training phase. Two methods for construction of the rules structure and various membership function types were investigated. It was observed that “fuzzy clustering” was preferable due to fewer rules and shorter training time. Also, Gaussian membership function showed better performance in comparison with triangular and trapezoidal shape membership functions. The number of input membership functions was optimized by trial and error process. Finally, the optimized ANFIS based on fuzzy clustering method was trained to unfold three spectra measured by other researchers which were 241Am-Be neutron source, high energy reference spectrum measured at PSI and fusion environment spectrum. These three spectra were not in the train set. Root Mean of Squared Errors (RMSEs) less than 0.026 and unfolded to original total fluence ratios near to one, were in agreement with the results reported by other researchers. The results showed that the ANFIS can be considered as a new and efficient method for fast neutron spectrum unfolding.