Unfolding of the Energy Spectrum of Neutron Flux with the "Random Forest" Machine Learning Algorithm

Abstract

Based on the "random forest" algorithm, a method has been developed for unfolding of the energy spectra of neutron flux from the readings of a Bonner multi-sphere spectrometer. To ensure smoothness of the spectrum function, the spectra were presented in the form of an expansion of Legendre polynomials. To improve the quality of the algorithm, four hyperparameters were optimized: the number of trees in the forest; maximum tree depth; the minimum number of samples required to split an internal node; the minimum number of samples required to be at a leaf node. It is shown that the hyperparameter “the minimum number of samples required to be at a leaf node” has the greatest impact on the quality of the algorithm’s work. The test sample shows that the unfolded spectra are close to the original ones and have a high correlation with them, equal to 0.78. The error calculated from the original and reconstructed spectra of the effective dose rate for isotropic irradiation does not exceed 44%. For global and local interpretation of the algorithm's results, the method of explainable artificial intelligence "local interpretable model-agnostic explanation (LIME)" was applied. It was shown that the model's performance is most influenced by measurements with moderator spheres of 5", 10" and 12" diameters.