Novel Multiple-point Geostatistical Algorithms Based On Graphics Processing And Model Evaluation

The characterization of geological structures is the foundation of reservoir research, which needs geologically realistic models. Currently, geostatistical simulation methodology is one of the main tools for reservoir simulation. Along with the technical improvement of reservoir exploration and production, numerical simulation is more and more important for field applications. Although the needs for high-resolution and large-scale simulation are increasing, the implementation of MPS methods is physical memory demanding, time consuming and parameters-sensitive, which limit the applications. The needs to release computational burden and improve geologically realistic interpretation are the main topics for MPS methods currently.This dissertation focuses on the similarities and intersections between MPS and computer graphics to improve MPS results and performance. A general-purpose graphic processor (GPGPU) based parallel scheme is implemented using Compute Unified Device Architecture (CUDA) to parallelism the calculation of high-order spatial statistics and significantly improve the performance. By introducing graph cuts algorithm which is a basic technique in computer graphics, a novel patch-based MPS approach is proposed, which has the ability to minimize pattern errors during generation as well as enables an iterative process to further modify the initial realization. Based on the reservoir models generated with the new approach, a series of parameter analysis, performance and results comparison are processed. Furthermore, hydrological models are performed to estimate the ability of geological pattern reconstruction. The main innovations including:1. A GPU-based parallel scheme is proposed for the calculation of geostatistical high-order statistics.a) Due to properties of equal distance for each lag of the high-order template, massive mathematical computing can easily parallelized on GPU using CUDA. By separating the calculation into two parts, that is, the calculation of spatial moments at each lag, and the combination of moments to cumulants, a two-stage parallel scheme is proposed and significantly improves the performance.b) To reduce the heavy memory demanding, an optimal memory management is proposed. By considering the transmit capacity of various kinds of memory, minimized data transferring strategies are organized.c) An adapted GPU reduction parallel algorithm is proposed to avoid the downgrade caused by atomic calculation. The sensitivity of thread structure is discussed to optimal this algorithm.d) Float point errors caused by CUDA compiler is discussed. By changing the sequence of calculation and data type, an improvement is achieved on calculation accuracy.2. A novel patch-based iterative approach is proposed for conditional simulation:a) A computer graphic technique known as graph cut is introduced to form a complete MPS algorithm.b) An iterative process is developed to continuously improve the simulation results by using the cut cost as a criterion.c) Achieve exact conditioning using the iterative process by taking the differences between realized valued and conditioning data at conditioning points as cut cost. Aterminal extension method is proposed to increase the local conditioning.3. Aseries results and performance analyses are processed.a) A concept of merging index is proposed to quantify the amount of verbatim copy of patch-based method.b) The sensitivity analysis of parameters proves that the proposed method is a converge method that not sensitive to parameters. This flexibility makes it user friendly especially for un-expert users.c) Comparisons with other MPS algorithms are generated using different methods based various statistical methods. The results show that the proposed method can generate at least as good realizations as classical MPS algorithms, whereas significantly improve the performance.4. Ground water transform models are processed on reservoir models to estimate the pattern reconstruction ability of MPS methods.