Improved seismic well tie by integrating variable-size window resampling with well-tie net

Abstract Accurate seismic well tie is essential for seismic inversion and reservoir characterization. The procedure of seismic well tie involves shifting, stretching and squeezing the synthetic seismogram computed from well logs to match the seismic traces at or near the borehole location. Numerous methods have been proposed for nonlinear alignment between synthetic and real seismograms. However, most well-tie methods are prone to over-stretching and the alignment result is sensitive to the chosen window size. To solve those problems, we propose a variable-size window resampling (VWR) algorithm and integrate with convolutional neural network (CNN) for automatic seismic well tie. Using VWR algorithm to reconstruct the waveforms in synthetic seismogram can simulate the variety of subsurface velocity. CNN can learn the characteristic of different waveforms and recognize the most correlated waveforms between synthetic and real seismograms for sequence alignment. We first use VWR algorithm to reconstruct a large number of synthetic seismograms for train set generation. We then build an CNN model that named well-tie net for training to learn the feature of different resampled synthetic seismograms. Finally, we use the well trained CNN model to segment the real seismogram and align with the synthetic seismogram for seismic well tie. We apply our method into the synthetic test and real seismic data with well logs and obtain high correlated seismic-well tie. We also compare with the conventional method dynamic time warping (DTW) to illustrate the effectiveness and robustness of our proposed method. Our proposed method can avoid the problem of over-stretching by using the variable-size window resampling algorithm and automatically tying the well to seismic trace using well-tie net. In addition, the train set for our method is generated automatically.