CN-121978758-A - Elastic wave field forward acceleration method and system based on parallel one-dimensional transverse convolution

CN121978758ACN 121978758 ACN121978758 ACN 121978758ACN-121978758-A

Abstract

The invention discloses an elastic wave field forward acceleration method and system based on parallel one-dimensional transverse convolution, wherein the method comprises the steps of S1, constructing a one-dimensional transverse convolution kernel according to a finite difference coefficient, S2, executing one-dimensional transverse convolution operation on an input matrix based on the one-dimensional transverse convolution kernel, calculating to obtain partial derivatives in the x direction, S3, converting data arrangement in the z direction into continuous data in the row direction through matrix transposition, executing one-dimensional transverse convolution operation on the basis of the transposed input matrix, calculating to obtain partial derivatives in the z direction, and S4, executing elastic wave field forward simulation based on the calculated partial derivatives in the x direction and the z direction. The technical scheme of the invention fully utilizes the high arithmetic strength and high calculation efficiency of convolution in the GPU, and solves the problem of slow data reading of the GPU by columns by using a matrix transposition technology.

Inventors

ZHANG YAN
CHEN BAIHAN
SUN YUHANG
YAO LIANGLIANG
FU BAOYU
TIAN FENG
LIU FANG

Assignees

东北石油大学

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (7)

1. The elastic wave field forward acceleration method based on parallel one-dimensional transverse convolution is characterized by comprising the following steps of: s1, constructing a one-dimensional transverse convolution kernel according to a finite difference coefficient; s2, based on the one-dimensional transverse convolution kernel, performing one-dimensional transverse convolution operation on an input matrix, and calculating to obtain a partial derivative in the x direction; S3, converting the data arrangement of the z direction into continuous data of the row direction through matrix transposition, and executing one-dimensional transverse convolution operation based on the transposed input matrix to obtain the partial derivative of the z direction through calculation; s4, performing forward modeling of the elastic wave field based on the calculated partial derivatives in the x direction and the z direction.
2. The elastic wave field forward acceleration method based on parallel one-dimensional transverse convolution according to claim 1, wherein said S1 comprises: Dividing the finite difference coefficient by the space step length delta x to obtain the weight of the one-dimensional transverse convolution kernel; setting the convolution kernel size to 1×n according to the finite difference order n; The weight of the one-dimensional transverse convolution kernel is a differential coefficient.
3. The elastic wave field forward acceleration method based on parallel one-dimensional transverse convolution according to claim 1, wherein said S2 comprises: Taking a one-dimensional transverse convolution kernel as a sliding window, and sliding from the upper left corner of the input matrix row by row; in each sliding process, multiplying the area covered by the convolution kernel with the convolution kernel weight element by element and summing; and taking the result of each calculation as a corresponding element of the partial derivative matrix to obtain a first partial derivative in the x direction.
4. The elastic wave field forward acceleration method based on parallel one-dimensional transverse convolution according to claim 1, wherein said S3 comprises: performing matrix transposition operation on the input matrix, and converting the data in the column direction into continuous arrangement in the row direction; performing one-dimensional transverse convolution operation on the transposed matrix to obtain an intermediate result; and performing matrix transposition operation on the intermediate result again to obtain a partial derivative in the z direction.
5. The elastic wave field forward acceleration method based on parallel one-dimensional transverse convolution according to claim 1, further comprising converting sliding calculation mode of the one-dimensional transverse convolution into parallel matrix multiplication to accelerate convolution calculation process.
6. The elastic wave field forward acceleration method based on parallel one-dimensional transversal convolution according to claim 5, wherein the step of converting the sliding calculation mode of the one-dimensional transversal convolution into parallel matrix multiplication comprises: setting the one-dimensional transverse convolution kernel as a row vector; Extracting a region covered by each step of sliding of convolution in an input matrix as a column vector; And carrying out parallel matrix multiplication calculation on the row vector and the column vectors to obtain a convolution result matrix.
7. An elastic wave field forward modeling acceleration system based on parallel one-dimensional transverse convolution, which is used for realizing the method of any one of claims 1-6, and is characterized by comprising a construction module, a first calculation module, a second calculation module and a forward modeling module; The construction module is used for constructing a one-dimensional transverse convolution kernel according to the finite difference coefficient; The first calculation module is used for performing one-dimensional transverse convolution operation on the input matrix based on the one-dimensional transverse convolution kernel, and calculating to obtain a partial derivative in the x direction; the second calculation module is used for converting the data arrangement in the z direction into continuous data in the row direction through matrix transposition, performing one-dimensional transverse convolution operation on the basis of the transposed input matrix, and calculating to obtain a partial derivative in the z direction; the forward model is used for executing forward modeling of the elastic wave field based on the calculated partial derivatives in the x direction and the z direction.

Description

Elastic wave field forward acceleration method and system based on parallel one-dimensional transverse convolution Technical Field The invention relates to the field of intersection of earth science and computer technology, in particular to an elastic wave field forward acceleration method and system based on parallel one-dimensional transverse convolution. Background In the field of oil and gas exploration and development, forward modeling is an important means for knowing the wave propagation process in underground medium, and provides scientific theoretical basis for searching underground oil and gas reservoirs. The time domain finite difference (FDTD) method has simple theory and accurate simulation, and is widely applied to forward numerical simulation. In two-dimensional forward numerical modeling, the FDTD method typically discretizes the continuous space into a number of uniformly fine grid points, with the wavefield update being performed on all grid points in each time step. This calculation process of time-stepping iteration-driven wave field evolution requires a large amount of calculation time. Especially when the sampling time span is longer, the number of required time iterations is increased, and the calculation time consumption is correspondingly increased. In addition, as the model is continuously scaled up and the finite difference order is gradually increased, the number of grid points needs to be increased in order to secure the imaging effect, which further increases the calculation time. The long calculation time greatly influences the real-time performance of forward modeling and is unfavorable for the development of subsequent work. Therefore, it is very important to study how to accelerate the calculation speed of the FDTD method. Currently, three methods for accelerating the FDTD calculation speed are mainly used. (1) FDTD calculation speed improvement methods based on different meshing schemes. Since the time-iterative computation of the wave field is an inherent property of the FDTD method, it is unavoidable. Different meshing schemes are used by students at home and abroad to reduce the number of grid points and the calculated amount. The reduction in the number of grid points means that the FDTD calculation accuracy is reduced. If it is still desired to maintain a higher accuracy of the calculation result, it is often necessary to use a higher order FDTD method or introduce other methods for improving accuracy. (2) The method divides a calculation domain into a plurality of subdomains by designing a parallel algorithm or calling an existing parallel calculation library, and distributes the subdomains to different processors (CPU or GPU) for processing so as to accelerate the calculation speed of each time step in the FDTD method. But this method requires a complicated boundary processing method to avoid boundary errors between subfields. (3) a FDTD calculation speed improvement method based on deep learning. The method utilizes a deep learning method to construct a model and learns the simulated wave field or the seismic data generated by the FDTD method. The trained model may enable fast prediction of wavefields or seismic data without the need for time iterative computations. Depth learning relies on data to implicitly learn physical laws, but network structures often lack explicit modeling of physical constraints such as wave equations, so that depth learning is generally weaker in wave field imaging accuracy than FDTD method wave field imaging. In view of the above analysis, the current elastic wave forward modeling has the main challenge that the accuracy of numerical simulation cannot be ensured while the calculation efficiency is improved. For this reason, it is necessary to design and develop an elastic wave forward modeling method and system that is consistent with the conventional finite difference method in terms of numerical simulation accuracy, but is more rapid in calculation. Disclosure of Invention In order to solve the technical problems in the background, the invention aims to provide an elastic wave field forward acceleration method and system based on parallel one-dimensional transverse convolution. Based on finite difference theory and convolution principle, finite difference coefficient is used as the kernel weight of one-dimensional transverse convolution operator, and derivative calculation of x-direction partial derivative in a first-order speed-stress formula is realized by using the one-dimensional transverse convolution operator. (2) By analyzing the GPU data reading principle and convolution calculation logic, the matrix transposition technology is used for exchanging the rows and columns of the physical quantity storage matrix, the data reading efficiency is improved, and the calculation of the derivative in the z direction is realized by using a one-dimensional transverse convolution operator. (3) The sliding calculation mode of the one-dimension