Supervised machine learning for discriminating fluid saturation and presence in subsurface reservoirs

Abstract

This paper describes a workflow that combines machine learning and frequency-related multi-attribute analyses to predict the occurrence of water, oil, and high or low gas saturations in three Miocene Gulf of Mexico clastic reservoirs, the Ursa, King Kong and Lisa Anne fields. The models are trained on labeled data from the Ursa Field dataset and validation is conducted on the King Kong and Lisa Anne Field datasets, comprising four wells with oil and varying gas saturations. We also analyze the impact of fluid-type and saturation on a suite of frequency-related attributes, including independent frequency magnitudes and spectral shape attributes.We also assess feature significance using SHAP Analysis and our results indicate that that high gas saturation is characterized by high magnitudes of low frequencies, increased bandwidth, and lower skewness and kurtosis values. Conversely, low gas saturation is distinguished by high magnitudes of medium frequencies, reduced bandwidth, and higher skewness and kurtosis values.Our workflow demonstrates the power and utility of supervised machine learning for reservoir fluid identification and saturation prediction. Human interpreters relying solely on post-stack seismic data may find the task of reservoir fluid predictions challenging and the results may be compromised by subjectivity. This study underscores the potential of machine learning to improve hydrocarbon exploration and production, providing a more robust approach to prospect risk assessment using post-stack seismic data and frequency attributes.