Backpropagation neural network approaches for assessing petrophysical properties of a carbonate reservoir in southeast Brazil

Abstract

This article applies backpropagation neural network algorithms to estimate petrophysical parameters, including lithology, porosity, permeability, and water saturation, in southeast Brazil’s post-salt Albian carbonate reservoir. Geophysical well logs (gamma ray, bulk density, neutron porosity, deep resistivity, and sonic), petrophysical laboratory data (porosity and permeability), and geological
information comprised the experimental datasets, while the water saturation was calculated with Archie’s equation. Firstly, rapid interpretation of field and laboratory data provided preliminary reservoir characterization, including boundaries and average porosity. Three backpropagation neural network algorithms were then implemented to assess the petrophysical parameters. During training, validation, and testing on reference hole LI10, Levenberg-Marquardt, Scaled Conjugate Gradient, and Bayesian Regularization techniques were used to fit the petrophysical parameters using the logs as input. Results demonstrate accurate and efficient processes, with high Pearson’s correlation coefficients and variable mean square error values across all parameters. Estimation
of lithology, water saturation, and porosity achieved high accuracy with all algorithms, while permeability posed the greatest challenge. Bayesian Regularization yielded the best performance, followed by Levenberg-Marquardt and Scaled Conjugate Gradient, though all produced reasonable estimates. Successful blind test at well LI03 confirms these techniques as promising approaches for
petrophysical characterization.