The Research Of Improved3-D Nonlinear Conjugate Gradient Inversion Of Magnetotelluric And Parallel Computing

Based on the MT3D nonlinear conjugate gradient algorithm and theprecondition method of2D NLCG inversion, we put forward an improved NLCGthree dimensional inversion algorithm. We extend the two-dimensional pre-regulatorto three-dimensional matrix, and use a3D matrix to build the preconditioning factor(not associated with the initial model). So the second derivative (Hessian matrix)calculation of objective function can be avoid, and the dependence of inversion andinitial model can be minimized. And two theoretical models are used to verify thecorrectness of the inversion procedure. The improved three-dimensionalmagnetotelluric inversion algorithm with parallel structure do not need big physicalmemory and high CPU speed of computer, so general PC is possible to use thisthree-dimensional inversion program. Nonlinear conjugate gradient inversionalgorithm is not necessary to store and calculate the Jacobian matrix, while calculatethe product of Jacobian matrix and any vector directly. Each iteration for inversion orlinear search with each frequency point only need one forwarding and five apparentforwarding, and it improves the efficiency.In addition, we use OPENMP technology to achieve the program with parallelcomputing, and it will assign the model calculation of each frequency point indifferent computer CPU, and make full use of existing computer’s CPU and memory.Moreover, this method is relatively independent on the initial model, therefore it doesnot affect the calculation precision in frequency parallel computing.Based on the theoretical synthetic data, the inversion algorithm is verified. Inaddition, through the forwarding and inversion of typical oil-bearing structure, wefound that the three-dimensional technology of magnetotelluric can play an importantrole in the oil and gas exploration work. Based on the measured data, we verify theeffectiveness of the inversion algorithms and procedures. Inversion results of Kayabedata show that the resistance of underground area is low upon500m. By contrast thedifferent pretreatment factor (associated with the initial model or not), the result of thefactor not related with initial model is better. And the three-dimensional model ismore close to the true underground structure. By comparing the three dimensionalinversion results of AMT data in Nihe with the results of controlled sourceelectromagnetic methods data, find that three-dimensional inversion can obtain betterresults while data’s SNR is bad. By the inversion results of the data in northwestUnited States of America, we find that upper crust and deep electric structure isinstead, and western area shows high resistivity in shallow and low resistivity in depth,and eastern area shows the opposite. The inversion results using different dimensionswith2D data in Xinjiang show that three-dimensional inversion can get better resultand its practicability is strong. Through the inversion result of Xinjiang data,3Dinversion give full consideration to the points of the horizontal direction as well as thefrequency of cases, so the result is more smooth and more accurate.