Neural networks based nuclear cross sections library

This work is concerned with the application of neural networks for the reproduction of nuclear cross sections, hence producing a neural network based nuclear cross section library as a mirror image to the classical nuclear cross section library such as the evaluated nuclear data file library. The advantage of this work is to present the same cross section spectrum with a limited data storage; namely; only the weights of the neural networks which are limited by the designed architecture of the neural network. Simple architecture implies small number of weights, and vise versa. The network, as it is trained with a selected set of data input ( energies of particles) and data output (cross sections), can predict the cross sections at any input energies lied in the range of the selected training data. The temperature dependence of cross sections can be easily handled either by training a multiple single input single output network structure; namely training number of networks with selected input/output data set of energies and cross sections at different temperatures, then storing the resulted weights of the networks, or alternatively, one designs a single multiple input single output network structure, where one trains a single network architecture with a set of two inputs data (temperature and energy) and the cross section output data. Of course, the resulted neural network based nuclear cross section library can be processed as raw data for significant applications in reactor physics. As an example; one can use collapsing codes to average cross sections and develop multigroup cross sections for reactor calculations. In this work, a selected data from the available nuclear cross sections for different elements and different types of reactions were used to train appropriate architecture of neural nets, and a reproduction of the trained data was achieved. Moreover, the nets were able to present the corresponding cross sections data for non trained data set. A user friendly interface has been developed for the prototype neural networks based nuclear cross sections library. This work demonstrates the innovative approach which could be recommended for sponsoring by the nuclear data section of the international atomic energy agency to develop the proposed neural network based nuclear cross section library. (author)