A supervised machine learning approach to discriminate reservoir fluid presence and saturation in the Gulf of Mexico using frequency and spectral shape attributes

Abstract

The first chapter in this research aims to define reliable attributes to differentiate subsurface fluids and measure their attenuation to provide insights into reservoir properties. The study investigates the effects of fluids and saturations on seismic data by analyzing a frequency-related suite of attributes using post-stack seismic field data from stacked reservoirs in Ursa Well #1 in the Gulf of Mexico. The results show that high attenuation reduces higher frequencies, resulting in a spectrum skewed towards lower frequencies with high kurtosis and low roughness, slope, and bandwidth. The attribute analysis presents that low amounts of gas saturation exhibit more attenuation than a fully saturated gas reservoir. Chapter 1 highlights the importance of spectral analysis, especially the spectrum's shape, in interpreting gas saturation and attenuation effects on post-stack seismic data allowing successful discrimination of water, oil, and high or low gas saturations. Furthermore, the second chapter presents a new workflow that applies supervised machine learning algorithms to predict the presence of hydrocarbon fluids and their economic viability using the frequency suite of seismic attributes. K-nearest neighbors, decision tree and random forest algorithms are tested on datasets for their robustness in classifying the fluid types. The SHAP values analysis is used to gather information on the importance of each attribute for the classification of each fluid class. The machine learning models are trained with the Ursa dataset and subsequentially predict in an expanded area around the training well. The King Kong and Lisa Anne fields, belonging to a different seismic survey, are used as a validation test for the machine learning models. The study concludes that the machine learning models, using the frequency suite of attributes, can predict water, oil, and high or low gas saturations within clastic reservoirs of the Miocene Gulf of Mexico. The decision tree and random forest models correctly predicted the fluid class in three out of four King Kong and Lisa Anne Field wells. The model predictions in the Ursa blind test expressed reliability in the models after observation of flat spots with correct density stacking orders. The promising results convey that using supervised machine learning algorithms for reservoir fluid identification and gas saturation predictions can revolutionize the field of hydrocarbon exploration and production by providing a more robust way to risk prospects using exclusively post-stack seismic data and frequency attributes.