Application of Geologically Constrained Machine Learning Method in Characterizing Paleokarst Reservoirs of Tarim Basin, China

Abstract

As deep carbonate fracture-cavity paleokarst reservoirs are deeply buried and highly heterogeneous, and the responded seismic signals have weak amplitudes and low signal-to-noise ratios. Machine learning in seismic exploration provides a new perspective to solve the above problems, which is rapidly developing with compelling results. Applying machine learning algorithms directly on deep seismic signals or seismic attributes of deep carbonate fracture-cavity reservoirs without any prior knowledge constraints will result in wasted computation and reduce the accuracy. We propose a method of combining geological constraints and machine learning to describe deep carbonate fracture-cavity paleokarst reservoirs. By empirical mode decomposition, the time–frequency features of the seismic data are obtained and then a sensitive frequency is selected using geological prior constraints, which is input to fuzzy C-means cluster for characterizing the reservoir distribution. Application on Tahe oilfield data shows the potential of highlighting subtle geologic structures that might otherwise escape unnoticed by applying machine learning directly.

Controls on Reservoir Heterogeneity of a Shallow-Marine Reservoir in Sawan Gas Field, SE Pakistan: Implications for Reservoir Quality Prediction Using Acoustic Impedance Inversion

“…The reservoir has high burial depth and geothermal gradient resulting in high temperature and pressure (170 • C and 37 MPa) [29]. The high burial conditions of depth and temperature cause the weakness and discontinuity in the seismic reflection signals due to which conventional methods cannot deliver better clarification of subsurface geology [30].…”