A deep residual convolutional neural network for automatic lithological facies identification in Brazilian pre-salt oilfield wellbore image logs

Field characterization in oil industry is a challenging task which aims to determine or estimate some of the petrophysical properties of a reservoir. These properties are expected to be later used in a non-deterministic workflow (which might depend on subjective interpretations carried out by each petrophysicist) whose aim is to answer questions such as whether exploiting the reservoir is feasible or not, how much hydrocarbon can be extracted from it and whether the geology of the reservoir will stand the stresses of exploitation without collapsing. Three properties stand-out amongst all the rest, as they are closely correlated to the amount of hydrocarbon present in the reservoir and also to the feasibility of its exploitation: permeability, porosity, and lithology. Obtaining reliable and robust measurements of these properties requires extracting core samples from the reservoir, which is, however, a very resource-consuming task. Hence, it is common to use other tools (such as wirelogging tools or borehole imagers) to extract several other properties from the reservoir, in an attempt to obtain information that might help experts to estimate these properties on those intervals at which no core sample could be extracted. In this context, estimating lithofacies helps petrophysicists to automatize the process of identifying the lithology of the reservoir. Previous work (Basu et al., 2002; Chai et al., 2009; Linek et al., 2007; Newberry et al., 2004; Hall et al., 1996) used a considerable amount of wirelog data and borehole image logs to induce the lithology of the reservoir. In the work presented in this paper, we introduce a new methodology for automatic reservoir lithofacies identification. Our model relies only on ultrasonic and microresistivity borehole image logs as inputs, which are characterized by a deep residual convolutional network which then infers the probability of each sample to be classified as each type of previously defined lithofacies classes. The method presented in this work was tested on a carbonate well from the Brazilian pre-salt oilfields, and it allowed us to obtain an average classification accuracy of 81.45% and an average area under the ROC curve of 92.70% for all classes, for the blind test sample.