Deep Carbonate Reservoir Hydrocarbon Detection Using Multiseismic Features Constrained Unsupervised Machine Learning

Abstract

Reservoir hydrocarbon detection is of great interest for reservoir characterization and quality assessment. However, deep carbonate reservoirs exhibit weak seismic response features, making it extremely difficult to extract and utilize reservoir information from seismic data, which leads to significant challenges for seismic-based reservoir detection techniques. The sparsity of the labeled samples often limits the application of supervised machine learning for seismic reservoir detection. This study proposes a multiseismic features constrained unsupervised machine learning approach for carbonate reservoir hydrocarbon detection in areas with few or no wells, which combines the multiple reservoir fluid feature extraction methods seismic-print analysis, high-resolution seismic attenuation gradient estimation, seismic dispersion analysis, and prestack simultaneous inversion, as well as advanced unsupervised machine learning isolation forest anomaly detection algorithm, to effectively extract and utilize the implicit reservoir pore-fluid information in seismic data. This method jointly uses multiple methods to extract multiseismic data features, which can overcome the problem that using a single method to extract seismic data features cannot fully reflect the reservoir pore-fluid information. Using unsupervised machine learning for multisource data feature fusion reservoir hydrocarbon detection can solve the problem that supervised machine learning's requirement for labeled data in deep-buried reservoir detection applications cannot be met. Actual field data application shows that the hydrocarbon detection results were consistent with the actual geologic understanding, which proves that the presented method is feasible and effective. This study provides a valuable insight and reference for reservoir detection in deep carbonate reservoirs with weak seismic responses.