Nonlinear Inversion Method Research Of Magnetotelluric Data

Megnetotelluric(MT)method,as an established geophysical technique,has been widely used for developing geotectology,crustal structure and deep mine resource prospecting.Geophysical inversion methods,as the bridge between geophysical data and the interpretation,always plays an important role in geophysics.In this thesis,we first brought the Fruit Fly Optimization Algorithm(FOA)in MT data inversion,with which we can avoid calculating the partial differential matrix and result rely on initial model and other weakness in linear optimization method.As a novel intelligent global optimization algorithm,FOA is inspired by the knowledge from the foraging behavior of fruit flies.The FOA is easy to understand and implement.From the analysis of classic FOA,we found that FOA is hard to deal with the high dimensional,multi-modal optimization problems and easy to trap in local minimum.We present an improved variant of classic FOA,called IFOA,to improve the FOA.We use the crossover operation and mutation operation from the differential evolution(DE)algorithm.Meanwhile,the mutation scale factor has been changed into the way that decrease gradually.The IFOA has been tested by several test functions.To analysis the optimal performance of IFOA,we compared the IFOA with FOA and differential evolution algorithm,the result show that the IFOA has the advantages of fast searching speed,high precision,excellent robustness and easy to escape from local minimum.After the functions test,IFOA has been applied into one dimensional MT data inversion.The average result shows it can search the optimal solution fast and deal with the data with different noise level.Another research field of this thesis is about the nonlinear Bayesian inversion.First summarizing the basic theory of Bayesian inversion and the posterior probability distribution(PDF)sampling methods.Bayesian inversion treats the model parameters as stochastic variables,and the inversion result is a PDF and easy to evaluate.Based on above,we utilize the Reversible jump Markov chain Monte Carlo sampling method for realizing MT trans-dimensional inversion.For the purpose of accelerating the convergence,the improved parallel tempering technique has been introduced into the inversion process.We test our improved trans-dimensional inversion on a three-layer model and a four-layers model to demonstrate the validity of the method,the result shows that the convergence of sampling process has been improved and the location of the layers also can be identified automatically.