A neuro fuzzy approach to recognize rock magnetic and lithological patterns in a stratigraphic well from the Llanos Foreland Basin (Colombia)

Abstract

The Neuro Fuzzy System (NFS) is a hybrid algorithm that combines fuzzy logic with neural networks. Since it can be used as a pattern recognition technique, we explore its potential to characterize the major lithological units encompassed by the first 512 m of the Colombian stratigraphic well Saltarin 1A (Guayabo and Leon Formations). Thus, we employ the NFS to infer the magnetic remanence S-ratio using bulk magnetic susceptibility (κ), κ-normalized saturation isothermal magnetization (SIRM κ) and/or volume of shale (Vsh) obtained from a gamma-ray log. The best results in terms of their corresponding Root Mean-Square Error (RMSE) values, throughout most of the upper Guayabo Formation, where magnetite seems to be an important magnetic phase, are attained with logκ and SIRM κ as input variables. Beyond 350 m downcore, the quality of the inference decreases over the Leon Formation, characterized by a significant presence of pyrrhotite. However, the extra input variable Vsh adjusts the inferred S-ratio to their experimental counterparts throughout this formation suggesting that the early diagenesis process that led to the formation of dispersed clay in these samples was also responsible for the formation of pyrrhotite. Hence the inclusion of manifold input data increases the ability of the net to predict S-ratio in complex geological settings with a sequence of changing lithologies, varying amounts and types of magnetic minerals, and different distributions of mineral grain sizes. In case these variables do not properly infer the actual S-ratio data, the extent of the different lithostratigraphic units would be still identifiable in some cases by the uneven quality of the correlation observed between inferred and experimental values.