Research On 3D Inversion Of Gravity Acceleration In CUDA Parallel Architecture

Gravity forward calculation is a method to solve the gravity anomaly of the observed surface on the ground surface. The inversion problem is to find the physical property and the shape of the source body according to the gravity anomaly value of the ground surface. Inversion of gravity anomaly plays a very important role in geophysical exploration, and it is the most convenient way to understand the characteristics of geologic body. Adopts cuboid mesh for the 3-D inversion technology, because the model parameters is easy to operate and can simulate arbitrary complex geological body; using a separate framework technology to reduce the storage capacity and other advantages has become a main technical means of gravity inversion. For a linear or nonlinear inverse scheme, even using the separation of grid technology, the forward calculation is still in the inversion process accounted for a large proportion, when the inversion has reached a certain scale, the inversion calculation 99% of the amount of calculation is determined by forward calculation produces, due to the accelerated forward calculation speed is of great significance for the inversion process.In this paper, the NVIDIA CUDA parallel computing platform is used to realize 3D gravity forward modeling and inversion calculation based on GPU. For forward calculation, based on the geometric trellis of equivalent storage technology, this dissertation put forward spatial locality of geometric trellis for the first time, and apply this property on the GPU forward calculation program for further optimization and achieved a four times faster than the ordinary version of the GPU forward calculation program. For inversion calculation, this dissertation clarify the inversion method based on the depth weighted is a strictly convex quadratic programming problem, and in the CUDA heterogeneous computing platform is designed and implemented based on the conjugate gradient method of the inversion process are given, also the numerical simulation results.