Application of an Artificial Neural Network ANN for the Correction of NMR Total Porosity, Effective Porosity and the Bulk Volume Irreducible Water Data Obtained from a Heavy Oil Reservoir using an Oil Based Mud Drilling Fluid in the Samaria: Tertiary on the Southeastern of Mexico

Abstract

Abstract Samaria Tertiary field in southeastern Mexico produces extra-heavy to medium grade oil (5 to 23° API) from sand-shale sequences. It is well known that the nuclear magnetic resonance (NMR) logging application can provide the total and effective porosity values, as well as free and related fluids independent of mineralogy. When the NMR logging tool is run in heavy oil deposits, oils of varying viscosities can be distinguished qualitatively. Unfortunately, however, in the presence of heavy oils, this measurement process underestimates the total porosity and effective porosity, and overestimates the bulk volume irreducible water (a key parameter for estimating permeability). Although the literature specifies methods that can correct this porosity and fluid volume bias by using fundamental magnetic resonance principles and conventional borehole information, they are only effective when the drilling fluid used is water-based mud. The artificial neural networks (ANN) have traditionally been used to correct well log data corrupted by tension pulls on the tool or degraded by poor borehole conditions. They have also been used for cases in which it is not feasible to obtain a traditional set of openhole well log data. In such cases, a neural network is used to generate synthetic openhole data from pulsed neutron logging data. In this application of ANN, the data is first manually interpreted to delineate uncorrupted zones of the data. This interpretation will correspond to zones in which there is no hydrocarbon or the hydrocarbon type is not heavy oil. The aggregate of such zones is used to define a training set. The remaining data is used to define the application data set. Traditional openhole data (such as resistivity, neutron, bulk density, and thorium) is used to train a neural network to predict NMR fluid volumes. The network obtained is then applied to the application data set, and it is shown that the resulting predicted NMR fluid volumes have been corrected for heavy oil effects in the presence of oil-based mud. This paper demonstrates the effectiveness of this approach and discusses the results obtained. It also demonstrates how the process can be automated.