Parallel Study Of Full Elastic Wave Inversion For Supercomputing Platform

Full waveform inversion of elastic wave is an important research method in seismic exploration.It can reconstruct high precision image from elastic parameters obtained from seismic data and provide basis for seismic exploration experts to carry out petroleum exploration.Compared with the traditional seismic exploration methods,elastic wave full waveform inversion has some advantages.On the one hand,elastic wave dynamic equation has more parameters than general acoustic equation,and it can describe the underground medium structure more accurately.On the other hand,complex geological structures can also be constructed more clearly.But this method comes at the cost of huge computing.With the rapid development of high performa-nce computing and the continuous innovation of the computing power of supercomputing platform,the large-scale implementation of this algorithm has gradually become possible.In this thesis,the elastic wave full waveform inversion algorithm is adopted.On the one hand,the finite difference method is used to carry out forward numerical simulation in the time domain,and on the other hand,the gradient calculation is simulated in the frequency domain to reduce the consumption of calculation cost in the time domain.This thesis studies the implementation of the parallelization algorithm of full waveform inversion,and the solution mainly includes the following parts:First,A method is proposed to improve the communication efficiency of full waveform inversion by using MPI repeated non-blocking communication.In order to solve the problem of too many times of initializing communication objects and too much overhead of repeated communication,the step size estimation of forward transmission,back transmission and the next iteration in full waveform inversion is accelerated by using the advantages of repeated non-blocking communication to initialize multiple communication at one time and the overlap between communication and computation,and the communication time is reduced.Then,In the iterative parallel process of full waveform inversion,the iterative calculation of the internal velocity v of the block boundary C-PML layer and the internal velocity v of the block are independent,so the iterative calculation of the internal velocity v of the block is completed by CPU,while the C-PML layer is calculated on GPU.At the same time,for blocks with multiple C-PML layers,the parallel optimization is carried out by using the overlap of computation and communication in CUDA multi-stream.Experiments verify the optimization effect of this scheme at a large scale.Finally,A simulation platform for full waveform inversion,which can also be used for full waveform inversion simulation on Web pages is developed,providing convenience for users to carry out simulation experiments.