Possibly the most underrated petrophysical parameter, the importance of water saturation cannot be emphasized enough with a whole range of petrophysical as well as reservoir engineering computations being dependent on its accurate determination leading to vital field development decisions; reserves estimation, waterflooding efficiency calculation and capillary pressure deduction. In 1942 Archie was first to present the equation to determine water saturation in a clean, non-clay reservoir. Ironically, ever since decades have passed with the intricacy of water saturation determination yet to be untangled in complex lithologies, especially in carbonates. Several researchers have tried to deconvolute the water distribution in composite formations by formulating empirical correlations that depend on log derived data which is not a very precise representation and hence no consensus exists among log analysts about which model can be universally used.
The use of computer generated algorithms, fuzzy logic and neural networking is picking up pace in the petroleum industry. Consequently, in this paper we show how Machine Learning can be used to generate a correlation, to determine water saturation in carbonate reservoirs, which is simple and practical to use in the sense that has less uncertainty in the parameters that it employs compared to existing models. In this work, multiple machine learning techniques namely, Artificial Neural Network (ANN), and Adaptive Neuro-Fuzzy Inference System (ANFIS) are used to estimate water saturation using conventional wireline log data as input parameters and the output is core dean-stark data. The data comprised of more than 2000 well log points which were reduced to around 150 corresponding to available core data. All the developed models are compared after a rigorous sensitivity analysis based on various artificial intelligence algorithms.
This work clearly shows that computer-based machine learning techniques can determine water saturation with a precision of approximately 94% when related to experimental core values. The developed correlation works extremely well in prediction mode with the shale affected log data as inputs. A comprehensive numerical and illustrative evaluation of the claimed accuracy is shown along with the error analysis between both the machine learning techniques used.