The necessary improvement of evaluated nuclear data for nuclear applications development is possible through new and high-quality experimental measurements. In particular, improving (n, n’) cross-section evaluations for faster neutrons than those involved in current reactors is a goal of interest for new reactor fuels. Our group at CNRS-IPHC has been running an experimental program to measure (n, n’ γ) cross-section using prompt γ-ray spectroscopy and neutron energy determination by time-of-flight, recording and analyzing data for 182,184,186W, 232Th, 233,235,238U. From the partial γ-transition measurements, the total (n, n’) cross-section has to be inferred, either by summing individual contributions (a method usually valid only up to a certain neutron energy), or by constraining reaction models with measured exclusive (n, n' γ) cross sections. This interpretation work is made difficult in (the usual) cases when not all the transitions going to the ground state could be measured. If that happens, one has to rely on filling the missing information by models or guess, reducing the accuracy of the final computed cross-section. Here we propose a new method, involving training a Neural Network on a calculated data set and using it to predict the (n, n’) cross-section from the experimental (n, n’ γ) ones. This allows a quick combination of models and experimental data. After detailing the method and checks performed for consistency, some test cases will be presented. Potential benefits, as well as the identified weakness, and future application will be discussed.