Development of a web application for an experiment to restore the neutron spectrum using neural network algorithms

The restoration of neutron spectra based on the measurement results by a multistep Banner spectrometer is an incorrectly stated inverse problem and requires special solution methods. The paper presents methods for spectrum reconstruction using a regression model of the random forest machine learning algorithm, as well as a neural network trained on synthetic data. The algorithms were trained and tested on a database consisting of 500,000 spectra artificially generated using the FRUIT method, and 340 real spectra from the IAEA collection and similar papers on the subject. The spectrometer readings for eight and ten retarder balls were used as input features of the model. It is shown that the developed algorithm is applicable for the reconstruction of neutron spectra. The reconstructed spectra are close to the original ones by the nature of the graph. The effective dose rate for isotropic radiation was calculated from the spectra, and it was shown that the average error in dose estimation is 25%.