An Accelerated Algorithm for 3D Inversion of Gravity Data Based on Improved Conjugate Gradient Method

Abstract

The 3D inversion algorithm for gravity data based on a smooth model constraint has been proven to yield a reasonable density distribution. However, as the amount of observed data and model parameters increases, the algorithm experiences issues with high memory consumption and prolonged computation time. Therefore, the corresponding problem in interpreting gravity inversion lies in developing a fast inversion algorithm. The conventional smooth model constraint inversion algorithm, based on regularization theory, requires the introduction of a model weighting function with a large matrix, and involves storage and operation of a large matrix with intermediate variables during inversion iteration, contributing significantly to the prolonged computation time. In this paper, a diagonal weight matrix is represented by vectorization, and the intermediate variable of the large matrix type in the iteration is replaced with the combination of a small matrix and a vector. Additionally, the algorithm flow of the conjugate gradient method is further optimized to minimize the number of vectors that need to be stored during iteration. As a result of these optimizations, the memory consumption of the algorithm during the operation process is successfully reduced. Finally, the experiments demonstrate the successful development of a fast 3D inversion algorithm for gravity data. Specifically, for a 80 × 80 × 20 mesh number inversion, our accelerated algorithm achieves an average speed of ~0.5 s per iteration, and the iterative process speeds up by a factor of 1000, providing an effective strategy for the fast inversion of large-scale data.