Subsurface Properties Determination From Nuclear Well-Logging Data Using Neural Networks

Determining subsurface properties through the use of surface or borehole data has been the subject of geophysical research for many years. In this study, we describe a fast and reliable innovative inversion approach, in conjunction with the application of neural networks and present the inversion results for three important subsurface parameters (porosity, salinity and oil saturation). The inversion procedure involves the minimisation of a functional, which embodies the neutron and 7-ray responses of the formation to a neutron source. The minimisation is achieved using an extensive global search of the search space. The forward problem is the transport of neutrons and photons within a sandstone formation and is investigated by simulating a typical nuclear welllogging tool. We obtained forward modelling results for both the steady-state and time-dependent problems, using a deterministic code, and utilised the predicted fluxes at the detector locations to train a set of neural networks. The deterministic code is based on the multi-group energy approximation, and solves the Boltzman transport equation for neutral particles using a variational discretisation procedure based on the even-parity form of the equation. The required cross-section data came from standard nuclear energy cross-section libraries and in particular the BUGLE96 library. A total of 67 energy groups were considered (47 neutrons and 20 gammas). For the steady-state problem, 67 neural networks, one for each energy group, were trained for the fluxes at the detector location. For the time-dependent problem, the time-domain signal was first decomposed into its corresponding Fourier coefficients and neural networks were set-up to train for each coefficient separately for every energy group. The inversion results we present are based on the steady-state solution.