Reservoir facies optimization using human-guided machine learning and probabilistic rock-physics templates

Reservoir facies are rock units that show distinctive features of rock types and fluid-flow properties. The characterization of reservoir facies helps us better understand the lifetime performance of reservoir. However, this process is complicated and challenging because we need to consider information from different sources with a wide range of scales (e.g., lab scale, well-log scale and seismic scale). This dissertation aims to characterize reservoir facies by integrating domain knowledge with statistical and machine learning techniques. To achieve this goal, I first present an active semi-supervised technique to classify well-log facies. This method integrates knowledge from domain experts and the data characteristics of wireline logs. It does not rely on training models or predefined classifiers. Instead, it utilizes small amounts of instructions from domain experts to guide facies clustering. The application on North Sea deep-water dataset illustrated that the semi-supervised approach could classify reservoir facies based on the data characteristics and simultaneously comply with the physical interpretations. In the second part of this dissertation, I developed a workflow using probabilistic rock-physics templates to determine the optimal reservoir facies models. Domain knowledge was imposed to build facies models with different pore-fluid parameters, and quantitative evaluations were performed to determine the optimal models. This workflow considers various types and magnitudes of errors from the petro-elastic, elastic and seismic domains. Use of Gassman equation and probabilistic rock-physics templates can mimic possible fluid types and different errors from the rock-physic modeling and well-log data. The Backus average accounts for resolution inconsistency from well-log to seismic scales. The low-frequency model building considers the possible misinterpretations of initial models in seismic inversion. I applied this workflow onto a dataset acquired in the Gulf of Mexico. Comparisons of different evaluation metrics suggested that scenario with five facies was the best facies model. Using the optimal facies models, I mapped the reservoir facies with different low-frequency models and determined areas with consistent estimations. This workflow outputs valuable information on the number and parameter distribution of facies models. It provides important guidelines for reservoir facies modeling when limited data are providedGeological Science