Machine Learning and Carbonate Reservoirs : A Production Data Perspective on Permeability Modelling

Occupying an important role in the global energy sector, carbonate reservoirs are the source of over half of the world’s remaining hydrocarbon reserves. Their successful recovery is key to maintaining a stable energy supply. In this context, an accurate modelling of the heterogeneous distribution of their petrophysical properties is essential for an optimized recovery of hydrocarbons. Among these petrophysical properties, permeability, which describes the ease of fluid flow in rocks, stands out as a vital factor for creating effective reservoir management and recovery strategies. Permeability measurements in carbonates are done using well core samples, which are time consuming and expensive to get, and since carbonate reservoirs in a field can extend laterally for kilometers, it is impossible to collect core samples from all wells in the field. Given the fact that a fraction of the total wells in the field has core samples, there is a high level of uncertainty on the permeability distribution in uncored wells and in the interwell region. Addressing this gap, this thesis aims to integrate machine learning models with production data to model permeability for the whole reservoir. In this study, 36 carbonates reservoir model incorporating different uncertainties on permeability distribution inherent to carbonate reservoirs, and their simulated production data are used to train and test an artificial neural network (ANN). These models are based on the COSTA model which represents a realistic set of heterogeneous carbonate reservoirs belonging to the Upper Kharaib member in UAE.

The ANN was rigorously tested, showing remarkable accuracy with the highest mean squared error (MSE) being 7.02 and the lowest MSE being 1.19, demonstrating its capability of modeling permeability. The ANN performance was further improved by integrating it with unsupervised learning techniques to cluster the production data of the 36 reservoir models into 3 different clusters based on the gas flow rates of production wells in each reservoir model. Consequently, separate neural networks were created for each cluster, each trained on reservoir models with a certain range and pattern of production data. This segmentation improved the ANN performance by 35%, highlighting the benefits of integrating unsupervised learning techniques in permeability modelling. One key finding in this thesis is understanding the effect of well interference on the ANN performance. It was noted that in cases where permeability increases from one well to another, the production data showed a decrease. Removing these producers from the training dataset led to an 10% decrease in ANN performance, highlighting their importance in capturing the non-linear relationship between permeability and production data.