Detector response unfolding using artificial neural networks

We present new results on the identification and unfolding of neutron spectra from the pulse height distribution measured with liquid scintillators. The novelty of the method consists of the dual use of linear and nonlinear artificial neural networks (ANNs). The linear networks solve the superposition problem in the general unfolding problem, whereas the nonlinear networks provide greater accuracy in the neutron source identification problem. Two additional new aspects of the present approach are (i) the use of a very accurate Monte Carlo code for the simulations needed in the training phase of the ANNs and (ii) the ability of the network to respond to short-time and therefore very noisy experimental measurements. This approach ensures sufficient accuracy, timeliness, and robustness to make it a candidate of choice for the heretofore unaddressed nuclear nonproliferation and safeguards applications in which both identification and unfolding are needed.