An Inversion Approach Combining Logging and Seismic Attributes Based on Machine Learning
Integrated reservoirs have been decreasing with many years of exploration and development.Thin reservoirs have gradually become the object of fine development.Logging technology has a high longitudinal resolution,because its depth sampling interval can be as small as 0.1 meter.But it can't ascertain the geological condition between the wells,because there is no way for the instrument to go down the well.Seismic exploration,the main means to search for oil and gas resources and has a good transverse continuity,is limited by the inability to distinguish thin layers less than a quarter of the length of seismic waves.The inversion combining logging and seismic data has the advantages of high longitudinal resolution of logging and good transverse continuity of seismic.The model-driven inversion method combining logging and seismic data requires a prior deterministic mapping operator,while the data-driven inversion method combining logging and seismic data does not requires an exact convolution model expression.The inversion mapping from seismic record to well logs does not need to establish an initial model,but only one information is seismic record in the input data which resulting in poor generalization ability.The two-dimensional matrix composed of logging attribute and seismic attribute is used as the input of the convolutional neural network to improve generalization ability.The inversion mapping from well logs to seismic record is not limited by convolution model,but depends on the initial model with multiple solutions.The result of inversion combining logging and seismic attributes is taken as the initial model,and a regularization constraint condition is added to the loss function to reduce the multiple solutions.The results of Marmousi-II model testing and the application of field data from Blackfoot area show that the inversion velocity of the proposed method is closer to the measured velocity,and the thin layer can be identified better from the predicted velocity profile.
velocity modelingjoint logging and seismic inversionattributesmachine learning
国家科技期刊平台
NSTL
万方数据
