Dynamic data integration into high resolution reservoir models using streamline-based inversion

A reservoir model derived from the static data only such as geologic, well, and seismic data, often fails to reproduce the past production history. The reservoir model then needs to be further conditioned to dynamic data such as pressure history, tracer response, and multiphase production history. Although it is fairly routine with modern geostatistical techniques to generate fine-scale reservoir models consisting of several hundred thousands of grid blocks, integration of dynamic data into such high-resolution models still remains a big challenge. Use of fast streamline-based simulation techniques can offer significant potential in this respect.;Streamline models can be advantageous in two ways. First, streamline simulators can serve as efficient forward model for history matching or the inverse modeling. Second, and more importantly, streamline models offer unique advantage in computing sensitivities of dynamic data with respect to reservoir parameters. Following an analogy between streamlines and seismic rays, the sensitivities can be formulated as integrals along streamlines. If the stationary streamline assumption is tolerable as in the cases of tracer transport or immiscible displacement with no significant total mobility changes, it is possible to compute these sensitivities analytically using a single simulation run. Application of the analytic sensitivity formulation to the estimation of NAPL (Non-Aqueous Phase Liquid) distribution in groundwater reveals three orders of magnitude faster inversion performance compared to the simulated annealing approach. For general flow situations where streamline updates are necessary, sensitivities are obtained by solving sensitivity equations along streamlines numerically.;Additional benefit from the seismic analogy is to utilize efficient seismic inversion techniques. Dynamic data integration is carried out in two steps: traveltime inversion followed by then the amplitude inversion. The two-step approach speeds up the inversion process and produces robust results from the quasi-linearity of traveltime formulation. Further improvement in computational efficiency is achieved from a multiscale approach that is based on hierarchical parameterization and scale-by-scale inversion. Also, the reduction of parameter space in the multiscale approach can avoid some adverse characteristics of inverse problems such as convergence to local minima, subjective choice of regularization constraints, and over-parameterization. The approaches proposed here are illustrated with several synthetic and field applications.