A Streamlined Neural Network Architecture for Magnetic Data Inversion

Data-driven methods based on deep learning have been applied to magnetic inversion and achieved excellent results. However, the existing neural network structures used for inversion are relatively complex, resulting in increased computational costs. Different from the inversion structure of existing encoder-decoder structures, this study designed a streamlined neural network inversion architecture based on the characteristics of magnetic anomaly forward modeling. The network structure only contains a decoder that maps magnetic anomaly data to a three-dimensional magnetic susceptibility model, which can save computational costs. First, the single-channel input data is transformed into multi-channel data through a transformation, then it is transformed into the dimensions of the magnetic susceptibility model through a four-layer decoder, and then the multi-channel data is transformed into a single channel through transformation, and finally the output is 3D magnetic susceptibility model. The transformation coefficients are trained by neural network. The neural network structure designed by this method is interpretable. It can reduce the parameters that need to be trained, reduce training time, and achieve high accuracy. It was verified through simulated and measured magnetic anomaly data, and high-precision inversion results were obtained. This idea can also be generalized to the inversion of other data.