As fast neutrons (~2 MeV) born from fission slow down (moderate) and thermalize (equilibrate), their de-Broglie wavelengths and energies approach the order of condensed matter inter-atomic spacing and collective, elementary system excitations (i.e, rotations, vibrations). The thermal scattering law (i.e, TSL, S(a,b) is a dynamical property that accounts for lattice contributions to neutron thermalization and impacts reaction rates, nuclear criticality and reactor core design calculations.

In this work, a Neural Thermal Scattering (NeTS) module (an artificial neural network) for nuclear graphite’s S(a,b) is trained to make low-latency, high-precision TSL predictions. PyTorch functionality is embedded within the Serpent Monte Carlo code, which is coupled to OpenFOAM. Additionally, a direct-NeTS method is developed in Serpent for generating continuous-temperature, continuous-energy, continuous-angle thermal scattering cross sections on-the-fly, directly from neural network S(a,b).

The NeTS-Serpent-OpenFOAM framework is tested for the steady-state and transient analysis of the TREAT M2 calibration experiment 2580, and shows experimentally-validated neutronic and thermal hydraulic improvements over existing methods. These methods support the improved design, characterization and deployment of advanced nuclear reactor systems.