Research On Semi-Airborne Transient Electromagnetic Inversion Algorithm

With the in-depth research on geophysics and the significant increase in computer calculation speed,semi-airborne transient electromagnetic detection technology has been widely used in the field of geophysical surveys.This method has the characteristics of high efficiency,low cost,and wide range.Semi-aircraft electromagnetic system detection equipment is also constantly being improved,and domestic research and development and integration of flight duration,high efficiency,multi-component and large sounding have been gradually realized in China.In recent years,the detection method carried by drones has gradually evolved into a development trend.While obtaining high-precision and high-quality measurement data,the higher the requirements for forward numerical simulation and inversion data interpretation.Electromagnetic inversion in geophysics has a history of several decades.The purpose of inversion is to use the electromagnetic response data of the receiving point to calculate the parameters of the medium model in turn based on the causal relationship between them.In recent years,various algorithms have been widely used in the interpretation of electromagnetic(EM)data in geophysics such as mineral,oil and gas and groundwater surveys.However,traditional inversion algorithms generally have the defects of low accuracy,slow speed and large amount of calculation.In order to overcome the above-mentioned problems,various new inversion algorithms have been continuously proposed.The combination of machine learning and traditional methods has become a research hotspot,which makes up for the shortcomings of traditional inversion methods and promotes the progress of inversion technology.This paper mainly studies the application of supervised descent method(SDM)in solving ground-based electrical source transient airborne electromagnetic systems.This is the first application of SDM method in the field of semi-aeronautical electromagnetics.SDM is a machine learning algorithm that contains two steps.The offline training phase learns the descending direction of each iteration step of one-dimensional(1-D)full-wave inversion(FWI)through the training data set with certain prior information,and then saves it.In the online prediction stage,the descent direction and data residuals are directly used to complete the full-wave inversion.In this article,we introduce the training process and prediction process of SDM,and give a formula description.Two different training methods are proposed to be suitable for inversion environments that cannot be used.This descent learning technology provides a new vision for combining traditional gradient inversion and machine learning-based inversion,as well as a kind of flexibility.The method integrates the prior information into the deterministic inversion through the physical model,which can avoid the local minimum problem and the complicated Jacobian matrix derivation problem,and achieve rapid convergence.The data verification results show that,with prior information,the SDM method is due to the traditional gradient inversion method both in terms of inversion speed and accuracy.In addition,the learning ability of this method has also been verified.