The Study On The Gravity And Magnetic Inversion Based On Parallel Computing And Deep Learning Algorithm

The technology of obtaining large-scale and high-quality gravity and magnetic data from the detection platform at a low cost has rapidly developed and gradually matured,and the supporting high-resolution data processing and interpretation technology is also constantly updated and improved.The 3D gravity and magnetic inversion can not only provide information about the geometric location of the horizontal boundary and buried depth of the anomaly source,but also provide quantitative information such as the size and distribution of physical parameters,so as to provide relevant references for subsequent geophysical exploration or geological interpretation.Therefore,the 3D inversion technology for modeling the density or magnetic susceptibility distribution of gravity and magnetic data is a vital part of data interpretation.For large-scale gravity and magnetic data,in theory,existing high-resolution inversion methods can be used.However,numerically,the computational complexity increases linearly with the size of the problem,which means that three-dimensional inversion faces two main obstacles.The first is the need for a large amount of computer memory to store the sensitivity matrix.Even the small-scale inversion of thousands of data points generally requires tens of thousands of units to form an underground three-dimensional model,which will exceed the available memory of the desktop.Then there is the CPU time required to multiply the dense matrix of the forward operator(the sensitivity matrix)by the data vector.These two obstacles directly limit the scale of what can actually be solved.This article analyzes and studies around these two obstacles.The forward calculation is the basis for implementing the inversion calculation.In each iteration of the inversion,the forward modeling needs to be applied—that is,the sensitivity matrix is applied to the prediction model—to calculate the prediction anomalies generated by the prediction model.We usually calculate and store the sensitivity matrix in advance for next use.But as the scale of the problem increases,the sensitivity matrix increases,and the calculation time increases.This paper proposes a sensitivity indexing technique based on an equivalent geometric grid,which only calculates and stores all possible values of the elements of the sensitivity matrix,and constructs a sensitivity matrix through the indexing technique.We verified the great advantages of this method in terms of reducing memory consumption through model tests,and at the same time reduced the time to calculate the sensitivity matrix to a certain extent,laying the foundation for large-scale data inversion.For fast calculation of large-scale data inversion,this paper proposes an inversion method based on parallel computing technology.MPI is used to implement CPU parallel calls to multiple GPUs,CUDA and OpenACC are used to implement GPU parallel calculations,and data transmission is reduced for parallel computing programs,Reduce memory access restrictions,hidden latency,and reduce instruction restrictions and other optimizations.We have proved through model tests that this method enables fast inversion of large-scale models.The introduction of dense data from the same underground model can more clearly depict the boundary position of anomalous bodies and outline their specific shapes.We also applied this method to the measured magnetic anomalies of the Jinchuan Copper Mine in Gansu.The inversion results are fast and stable,which is consistent with the geological data.The application of the concept of deep learning to the automatic interpretation of geophysical data is a research hotspot today,but the field of using deep learning methods to perform gravity and magnetic inversion to restore the distribution of underground physical properties is still blank.This paper realizes the inversion of gravity and magnetic data based on convolutional neural network,and designs a relatively simple and effective convolutional neural network,which greatly reduces the storage and training parameters required for gravity data density imaging,and can Efficient and accurate physical property imaging of gravity and magnetic data.The algorithm was verified by model tests,and the network was applied to the geomagnetic interpretation of aeromagnetic anomalies in the Jinchuan mining area.