Fast Electromagnetic Inversion Algorithm Based On Machine Learning

Electromagnetic(EM)inversion techniques and their applications have developed rapidly in the past decades thanks to the fast increase of computer speed and memory and the development of fast forward modeling methods.The general purpose of EM inversion is to retrieve the model parameters such as shapes,locations or constitutive parameters of unknown objects in a specific region by using the scattered field data measured at the receiver arrays located outside that regioin.However,due to the shortcomings of the traditional inversion algorithm,such as slow inversion speed and low accuracy,the further development and application of the technology are limited.Also thanks to the rapid development of hardware devices,machine learning has become a research hotspot in various fields in recent years.In the field of electromagnetic inversion,machine learning also promotes the further development of electromagnetic inversion technology.Fast or high-precision electromagnetic inversion methods based on machine learning are constantly proposed,which make up for the defects of traditional technology.In this work,machine learning is combined with traditional electromagnetic inversion technology to speed up the inversion speed and improve the inversion accuracy.In this work,two algorithms are proposed.One is the 3-D full-wave inversion method which combines the 3-D Markov random field(MRF)with the traditional inversion method,variational Born iterative method(VBIM).In each iteration,VBIM first reconstructs the model parameters of all discretized cells in the inversion domain by solving the discretized data equations.Then,MRF is adopted to classify the cells according to the reconstructed parameter values,i.e.to determine whether a certain cell belongs to the background or the scatterer.At last,partial cells classified as "background"are removed and partial cells classified as "scatterer" are merged.Consequently,the unknowns in the following VBIM iterations are reduced.In the iterative process,due to the continuous reduction of unknowns,the computational cost is constantly reduced,so as to speed up the inversion speed.Another function of the reduction of unknowns is to alleviate the under-determinedness and ill-posedness of the discretized data equations,so as to obtain more accurate solutions.The other is the 2-D full-wave inversion method which combines linear sampling method(LSM),convolutional neural network(CNN)and traditional inversion method,Born iterative method(BIM).First,the LSM qualitatively reconstructs the preliminary shapes and locations of the scatterers.Then,the CNN is used to further refine the shapes of the scatterers.At last,the Born iterative method(BIM)is implemented to quantitatively invert for the permittivity and conductivity of scatterers inside the downsized inversion domain.The two algorithms in this work are verified by corresponding numerical examples.The results show that the method in this work is feasible and effective,and compared with the traditional inversion method,it can achieve higher reconstruction accuracy and lower computational cost,and also has anti-noise capability.Furthermore,laboratory experimental data are also used to verify the feasibility of the second algorithm.