Gravity measurements have a wide range of applications in the detection of natural resources, mapping bedrock topography, and microgravity research (Nabighian et al., 2005). From a geophysical perspective, the main outcome when interpreting gravity data is the determination of the geometric and physical characteristics of rock bodies.It is generally assumed that the subsurface reconstruction that is obtained from observed data is composed of a mass of discrete rectangular blocks with contrasting densities (Boulanger & Chouteau, 2001). Further, due to noise in field data, the application of inverting gravity data is limited, and gravity sources are inherently non-unique. In recent years, several approaches have been proposed to obtain a reasonable inversion solution, most of which have been employed for sparse inversion (Liu et al., 2018). Indeed, researchers expect to develop a sparse solution, Here, most elements of density vectors become zero, and a few remain large.Machine learning (ML) is a branch of artificial intelligence that has produced good results when applied to prediction systems, fraud alerts, image recognition, and spam filters (Bobadilla et al., 2013;Ravisankar et al., 2011). The main reason for this success is that it allows computer systems to "learn" by driving a large number of data pairs for training. In the field of geophysics, the application of nonlinear techniques in ML can be traced back to the 1990s. Specifically, Guan et al. (1998) combined a neural network with the inverted theory of gravity and magnetic anomalies, thereby developing a pseudo back propagation (BP) neural network method that can invert gravity and magnetic data. Although this method does not have a sample set to update the network weights, it loses the ability of self-learning. Guo et al. ( 2012) investigated a 3D gravity inversion method with a large number of parameters based on a BP neural network. Compared with the BP neural network, the radial basis function (RBF) provides the additional advantages of nonlinear mapping under constraint conditions and good generalization capabilities. On this basis, Geng and Yang (2013)