Unsupervised Learning for Automated Seismic-Well Tie

Abstract

Summary As an essential process in subsurface interpretation, seismic-well tie aims at calibrating the depth-domain well measurements with the time-domain seismic records for reliable reservoir calibration and modeling. Such a process usually requires plenty of manual efforts, primarily stretching and/or squeezing synthetic seismogram to match actual seismic traces. In this paper, we propose an unsupervised workflow for automated seismic-well tie that addresses the issue of lacking manually prepared training labels in most real projects. It starts with preparing training data by generating 1D time-shift curves, warping the original less accurate time-depth relationships (TDRs) and synthesizing the corresponding seismograms based on the convolutional model at multiple target wells. With the paired seismograms from original and warped TDRs, the next step is to train a 1D FlowNet that estimates the time shifts applied for TDR warping. At the stage of inference, by feeding the synthetic seismogram from the original TDR and the observed seismic trace at a well, the trained FlowNet predicts the time shifts necessary for revising the TDR that would lead to optimal tying. The proposed workflow is tested on the F3 Netherlands dataset, and the results demonstrate improved tying quality over the traditional methods such as dynamic time warping.