24 May 2022 in Geophysics
Three dimensional gravity inversion is an effective way to extract subsurface density distribution from gravity data. Different from the conventional geophysics-based inversions, machine-learning-based inversion is a data-driven method mapping the observed data to a 3D model. We have developed a new machine-learning-based inversion method by establishing a decomposition network (DecNet). Unlike existing machine-learning-based inversion methods, the proposed DecNet method is a mapping from 2D to 2D, which requires less training time and memory space. Instead of learning the density information of each grid point, this network learns the boundary position, vertical center, thickness, and density distribution by 2D-to-2D mapping and reconstructs the 3D model by using these predicted parameters. Furthermore, by using the highly accurate boundary information learned from this network as supplement information, the DecNet method is optimized into a DecNetB method. By comparing the least-squares inversion and U-Net inversion on synthetic and real survey data, the DecNet and DecNetB methods have shown the advantage in dealing with inverse problems for targets with boundaries.
1 June 2019 in Computers & Geosciences
4 January 2023 in Geophysics
A novel approach is developed to implement the forward modeling of a potential field caused by prismatic grids on gridded observations with higher efficiency and less memory. Based on the spherical symmetry of gravitational potential, we summarize the parity, symmetry, and interconversion relations from derivatives and higher derivatives of gravitational potential, which reveal that there are vast repetitive calculation and storage redundancies in conventional forward modeling of a potential field for prismatic grids, especially for cubic grids. These properties help to reduce not only the size of a single kernel matrix but also the number of necessary kernels in multicomponent forward modeling of a potential field, such as a gradient tensor field or joint inversion of a potential field. Several experiments on the synthetic and the realistic Bishop models are presented to illustrate how these properties are used in the forward modeling of gravity and magnetic gradient tensors. Based on the proposed properties, a 2D fast Fourier transform is performed to accelerate the discrete deconvolution of the kernel and the model parameters. With the new method, one single kernel can be reduced to half the number of model parameters, and the memory requirement and time cost of our method can be reduced by 1/8 and 1/2, respectively, compared with previous works. In forward modeling of six gravity gradient tensor components, only two kernels need to be computed and stored, which are 1/3 of the conventional method. These examples indicate that, with high precision, our method requires less memory and less time than conventional methods. It shows potential in the fast large-scale inversion of gravity, magnetic, and gradient tensors with limited hardware resources.
13 April 2026 in Near Surface Geophysics