Thermal neutron beam optimization for PGNAA applications using Q-learning algorithm and neural network

As a powerful, non-destructive analysis tool based on thermal neutron capture reaction, prompt gamma neutron activation analysis (PGNAA) indeed requires the appropriate neutron source. Neutrons produced by electron Linac-based neutron sources should be thermalized to be appropriate for PGNAA. As a result, thermalization devices (TDs) are used for the usual fast neutron beam to simultaneously maximize the thermal neutron flux and minimize the non- thermal neutron flux at the beam port of TD. To achieve the desired thermal neutron flux, the optimized geometry of TD including the proper materials for moderators and collimator, as well as the optimized dimensions are required. In this context, TD optimization using only Monte Carlo approaches such as MCNP is a multi-parameter problem and time-consuming task. In this work, multilayer perceptron (MLP) neural network has been applied in combination with Q-learning algorithm to optimize the geometry of TD containing collimator and two moderators. Using MLP, both thickness and diameter of the collimator at the beam port of TD have first been optimized for different input electron energies of Linac as well as for moderators’ thickness values and the collimator. Then, the MLP has been learned by the thermal and non-thermal neutron flux simultaneously at the beam port of TD calculated by MCNPX2.6 code. After selecting the optimized geometry of the collimator, a combination of Q-learning algorithm and MLP artificial neural network have been used to find the optimal moderators’ thickness for different input electron energies of Linac. Results verify that the final optimum setup can be obtained based on the prepared dataset in a considerably smaller number of simulations compared to conventional calculation methods as implemented in MCNP.