Machine learning assisted low-frequency model building for AVO inversion
Master thesis
Permanent lenke
http://hdl.handle.net/11250/2561830Utgivelsesdato
2018Metadata
Vis full innførselSamlinger
Sammendrag
AVO inversion is a valuable tool to estimate absolute reservoir properties from prestack seismic data. The bandwidth limitation of the seismic requires that the missing low frequencies must be added to the inversion in order to estimate absolute values. The standard method is to build a low-frequency model by interpolating and extrapolating low-pass filtered well logs guided by interpreted horizons, which often leads to artifacts and has no geologic insight. Artifacts and uncertainties in the low-frequency model manifest themselves in the inverted estimates, leading to incorrect interpretations and bad business decisions. In this study, an innovative methodology is proposed based on machine learning to find the non-linear relationship between seismic attributes and logged elastic properties from available wells, then the relation is used to estimate the low-frequency content of a target property away from well control. Only relevant attributes with a justifiable physical relationship with target elastic properties were used, such as AVO attributes and relative impedance. The methodology was applied in a practical exploration case and validated for a true blind well location. Even though the statistical basis of the method becomes less robust in areas with very limited well availability, the results indicate that this methodology is able to estimate more accurately low-frequency content of elastic properties than the conventional method.