Research And Application Of The Parallel Inversion Algorithms Based On The Full Tensor Gravity Gradiometry Data

Full tensor gravity gradiometry(FTG) data is a kind of high-precision data which has been widely applied in the resource and geological explorations in the recent years. Compared with traditional gravity anomaly data, FTG includes more information and its signal-to-noise ratio is higher. Especially as the depth of the exploration for crust keeps increasing, it is more difficult for the exploration. The higher precision and larger amount of the data is required. The underground information should be detected by more advanced processing and interpretation methods and approaches. To take full advantage of the high precision of FTG data and solve the problems of the long-lasting computing and massive memory resource brought by the geophysical inversion with large data, the computer science and geoscience need to be combined and the research should be carried from the point of applying parallel computing.Geophysical inversion is an interpretation method computing and solving the geological structure with the geophysical field data recorded in the observation surface. The physical properties included in FTG data could be reflected by the inversion; for larger data, the parallel computing is used for improving computing speed and solving the problem of memory usage, making the inversion implemented. Hence, the effect of the parallel algorithm depends on the effectiveness of the inversion algorithm. This thesis is based on the principle of the inversion with FTG data, aiming at solving the problems above in the algorithms. Many parallel algorithms and programs are researched and designed, the desired results are achieved.First, it is started from the forwarding theory of the gravity anomaly and its gradiometry data generated from the geological bodies. The way dividing into prism cells is chosen to discretize the forwarding equations, which is the basis of the inversion. Reweighted focusing inversion and the inversion of probability tomography are two common algorithms. The research on these two algorithms are used with single type of data at present and the space resolution of their results needs improvement. In the calculations, there exist some problems that lots of memory is consumed by the forwarding matrix, the iteration stopping is hard to control and so on. In this thesis, it is proposed that FTG data is used for reweighted focusing inversion and three tensors of FTG data are used for the joint inversion of probability tomography. The algorithms are improved to increase the uniqueness of the solution making the improved density distribution closer to the real underground situation. The improvements include: prior density constraints, multiple-component joint inversion, the coefficient of determination as iteration stopping condition, recalculating the density variation and using depth weighting matrix for the inversion of probability tomography. It is found in the theoretical model tests that reweighted focusing inversion with FTG data and the inversion of probability tomography with three tensors have a higher space resolution than the other algorithms, both of which have the advantage on the vertical and horizontal direction, respectively. The improved algorithms could recognize the geological boundary more accurately and divide the adjacent geological bodies. The iterations could be stopped when the desired results are obtained with the coefficient of determination that improves the numerical precision of the results. In the real data test, the real airborne FTG data of Vinton Dome in the U.S.A. is used and the results are consistent with the other scholars’ research, meet the seismic and logging information. It is proved in the above tests that two improved inversion algorithms are feasible and reliable, both of them have the certain anti-noise ability.Then the research on the basis of parallel computing technology is considered. The parallel computer architectures, parallel programming models and languages involved in the thesis are stated. That provides a theoretical reference for the parallel algorithm design. Message Passing Interface(MPI) and Compute Unified Device Architecture(CUDA) are two main parallel pattern at present which are widely applied in the parallel geophysical inversion algorithms. The existing parallel algorithms of reweighted focusing inversion and inversion of probability tomography lack the research of joint inversion. Single GPU is always used only for parallel computing that limits the acceleration and processed data amount. According to the analysis on the inversion, the time-consuming parts in the calculations include the matrix element calculation and matrix-vector multiplication and the iterations in the inversion increase the proportion in total running time for these parts. In the past research and application, the consideration of the memory usage caused by the large forwarding matrix and memory optimization in the iterative computation is uncommon, the performance analysis is few.Aiming at solving these problems, the thesis implements parallel computing for reweighted focusing inversion with MPI, taking advantage of multiple-node memory to solve the problem of memory consumption caused by forwarding matrix, the good parallel efficiency and scalability are obtained and the maximum speedup is far more than 100. The calculated amount for inversion of probability tomography is less than the former, the method of instant computing is used for the elements in forwarding matrix, CUDA is adopted for parallel computing based on single GPU and multiple GPUs, respectively, and the good acceleration has been obtained, the fastest running time is less than 30 s when the data size is 100 ?100 ?20. It is proved that the algorithm with multiple GPUs has the stronger acceleration ability and can process larger data than that with single GPU. The parallel algorithms use the way of dividing rows or columns, which is combined with improved inversion algorithms, for matrix-vector multiplication and matrix element calculation. That reduces unnecessary communication overhead among the processes or data transfer between host and GPU memory. In the performance tests, the performance metric is analyzed and the related formulas are deducted according to the features of the parallel inversion algorithms. That improves the accuracy of the analysis and regulates the methods for analysis.To take more advantages of the computer performance and implement different algorithms on the corresponding computer hardware, and to overcome the acceleration performance obstacles when using single pattern for the parallel program, combining different parallel methods for hybrid programming should be the main way to solve the inversion problems of large-scale data. In this thesis, MPI, CUDA and Open Multi-Processing(Open MP) are selected for hybrid programming, including the parallel programs of the inversion of probability tomography based on MPI-Open MP, MPI-CUDA and Open MP-CUDA. It is demonstrated in the performance tests that the hybrid programs above perform well on the acceleration and help to make a further improvement for the inversion with larger data. All parallel algorithms in the thesis are summarized and compared with each other. That provides the reference for the future research and application. At last, MPI program of reweighted focusing inversion and Open MP-CUDA program of the inversion of probability tomography are chosen to construct a hybrid parallel inversion system according to the features of the inversion algorithms and their parallel programs. This system uses the graphics processing node and ordinary computing nodes in the cluster as the platform which could implement automated data input and output, and parallel computing. The system can also choose the algorithm for the input data intelligently taking full advantage of high-precision data and the space resolution of improved inversion algorithms.In conclusion, the inversion algorithms with FTG data are improved in this thesis for a better density distribution and higher space resolution. That provides an accurate algorithm basis of the parallel computing. Different parallel algorithms and their programs are designed which have a good accelerating ability and could solve the problems caused by massive data. That meets the need of the processing and interpretation of the big high-precision data in the deep exploration. It is confirmed in the tests and analysis that the improved inversion algorithms and the parallel algorithms are effective, and have the theory significance and value of practical application. This thesis provides the important reference support for the future research.