Well-log validated waveform inversion of reflection seismic data
Sergio J. Romahn
This thesis develops, examines and refines the basic waveform inversion procedures of reflection seismic data under an alternative perspective that makes an effort to assign to each element of full-waveform inversion (FWI) workflow an element of standard seismic processing. The basic procedures can be summarized as an iterative cycle of modelling, migration and inversion. The modelling part was carried out with a fourth order finite-difference scalar acoustic algorithm, assuming a 2D medium governed by the constant density acoustic wave equation. The two-way wave operators (analogous to reverse-time migration), used as standard engine in FWI to produce the gradient, is replaced with one-way wave operators. The role of the pseudo-Hessian as a gradient preconditioner is substituted with a deconvolution imaging condition. A key element of this methodology is the incorporation of log information to calibrate the update direction (i.e., a well-log validation approach), rather than determining a scalar step-length with a line search (i.e., a data-validation approach). This inversion perspective leads to remarkably accurate results and significant computational savings. Once a well-log validated waveform procedure was set up, I evaluated the scope of well-log validation under several factors (such as geological complexity, well location, log interval, and well-data uncertainty) that may influence the performance of this technique due to the welllog limitation of providing punctual information. I also analyzed the repercussions that the seismic acquisition parameters, the random noise and a deficient initial model would have on the inversion. Having established pros and cons of well-log validated waveform inversion with one-way wave equation migration, I addressed the unknown-wavelet issue. A process of updating both the amplitude and phase of an initial wavelet estimate is set up to produce stable inversions in both synthetic and field environments. In the latter case, profiles extracted from a log-validated waveform inversion of the Hussar land data set, when compared to two blind validation wells, are observed to accurately reproduce structures in zones producing significant reflection energy.