Seismic-constrained Reservoir Characterization And Uncertainty Quantitative Evaluation

Obtaining accurate reservoir parameters(e.g.,elastic parameters,petrophysical parameters,lithology/fluid)is an important part of reservoir characterization.Seismic data is the main guiding information for reservoir characterization,and plays an important role for predicting of elastic parameters,petrophysical parameters,and lithology/fluid.However,inverting seismic data to reservoir parameters has several challenges due to limited data bandwidth,noise,approximation of the physical model,and incomplete data coverage.Therefore,this dissertation proposes corresponding solutions to the problems associated with the model-driven and data-driven methods,starting from improving the accuracy of reservoir characterization.The validity of the proposed methodology has also been demonstrated by numerical examples of the model and actual data.Seismic inversion(pre-,post-stack)is a common method for hydrocarbon reservoir characterization,as it consists of a proven and effective approach to derive reservoir properties from reflectivity seismic data.In a linear situation,the traditional multi-step inversion method based on elastic properties ignores error propagation effects,lateral correlation of reservoir and the limited resolution of inversion results,which increases the risk of reservoir evaluation.A joint estimation of elastic,petrophysical properties and lithology/fluid based on Bayesian seismic inversion with geological structural constraint is proposed to improve reservoir characterization accuracy.However,in some complex geological regions,the use of linear operators often leads to low accuracy of prediction results.The nonlinear operator can reduce the computational errors brought by the linear approximation and improve the accuracy of inversion results.In the nonlinear case,a geological structure guided Markov chain Monte Carlo(MCMC)strategy is implemented to realize the probability density function characterization of reservoir parameters,taking into account the time cost and memory consumption.The geological structure information obtained using plane wave destruction(PWD)is incorporated to the MCMC based inversion algorithm in the form of dips yields more geologically meaningful results.The hybrid MCMC and BLI strategy,which takes advantage of BLI and the block coordinate descent(BCD)algorithm to reduce memory consumption and time cost.However,the existing physical models are only approximate representations of the reservoir parameters and seismic data,and it is difficult to find a mathematical model that can be used to fully and accurately characterize the relationship between them.In order to characterize the relationship more accurately,this paper proposes a neural network framework with a priori initial model constraints to construct matching functions between reservoir parameters and seismic data directly using training data to achieve stable seismic inversion and its uncertainty evaluation.Prediction of lithology/fluid characteristics is always the bottleneck problem and difficulty of reservoir characterization.However,the approximate physical model,the illposed nature of the inverse problem,and the cumulative error of the multi-step process often leads to low accuracy of the prediction results of lithology/fluid.To overcome these issues,I proposed a spatially coupled data-driven(convolution neural network,CNN)approach for lithology/fluid prediction from post-stack seismic data and well observation.The LFP results of the proposed approach are more laterally continuous and geologically reliable than the LFP results of the point-by-point.I determined the effectiveness of this methodology on a 2D synthetic model and a 3D field seismic data set.