Time-lapse seismic imaging, full-waveform inversion, and uncertainty quantification

Xin Fu

Time-lapse seismic, also known as 4D seismic, is a powerful tool for monitoring subsurface changes over time. By comparing seismic data acquired at different intervals, it enables the detection and characterization of dynamic reservoir processes, aiding in reservoir management, production, optimization, and enhanced oil recovery. It has applications in geothermal energy, CO2 storage monitoring, and environmental impact assessment. However, accurate analysis of time-lapse seismic data remains a challenging task. It requires well-repeated time-lapse seismic surveys, including well-repeated acquisition geometry and equipment as well as well-repeated ambient noise. This thesis is to alleviate the non-repeatability issues in time-lapse seismic imaging and full-waveform inversion (FWI), and to realize the uncertain quantification for time-lapse seismic waveform inversion. A time-lapse imaging approach that involves two new frequency-domain matching filters is developed. The first filter requires source wavelet estimates from both baseline and monitoring data, while the second filter is source-independent but more sensitive to data noise. By applying these filters, we successfully reduce source wavelet non-repeatability, and the new approach improves the accuracy of time-lapse imaging. Furthermore, a stepsize-sharing time-lapse FWI strategy is designed to reduce artifacts caused by the variability of convergence in conventional strategies. The strategy demonstrates good adaptivity in different tested realistic scenarios using synthetic data. It is stable for scenarios using biased starting models, while the conventional strategies fail in this regard. Moreover, to realize the uncertain quantification, a Bayesian time-lapse FWI procedure, based on a Markov chain Monte Carlo (MCMC) algorithm, is formulated.The formulation employs several existing strategies, including the use of a double-difference time-lapse FWI, incorporation of time-domain multi-source data, and application of a local updating target-oriented inversion. It incorporates these within a stochastic framework,involving the computation of model covariance with an adaptive Metropolis algorithm, and a method to estimate data error statistics based on the features of time-lapse difference data is incorporated. A random walk Metropolis-Hastings MCMC is adopted for optimization.