Full-Current RMT Parallel Calculation And Preliminary Study On PINN Numerical Simulation

Radio-magnetotelluric method(RMT)is one of the important tools for shallow surface exploration.Its detection frequency is between 10 k Hz and 1 MHz,and it can carry out frequency domain electromagnetic sounding in the shallow hundreds of meters.The traditional RMT data forward and inversion often use the Magnetotelluric(MT)program for quasi-static hypothesis processing,ignoring the influence of displacement current.However,simulation studies have shown that the influence of displacement current cannot be ignored in high resistivity and high dielectric constant media.Therefore,the RMT exploration frequency must adopt the full current condition and develop a forward and inversion program suitable for RMT.In this paper,the full current RMT forward and inversion problem is studied and explored in parallel computing and intelligent forward and inversion.The main work and conclusions include: 1)The traditional serial algorithm deals with multi-frequency point calculation,which has the disadvantages of long time.In order to solve this problem,a parallel acceleration forward algorithm based on Julia is proposed.Julia ’s distributed parallel is used to assign the calculation of frequency points to different processes for solving,so as to achieve the purpose of acceleration.The two-dimensional RMT problem is discretized by unstructured finite element to improve the accuracy of the solution.By calculating several high-resistance and high-dielectric models,the influence of displacement current on the apparent resistivity and phase of RMT exploration frequency band is analyzed.The results show that when there is a high-resistance covering layer in the shallow part,the apparent resistivity and phase response of RMT calculated based on quasi-static assumptions are high,and the higher the frequency,the greater the resistivity of the covering layer,the greater the response deviation.The comparison of two parallel examples of different scales proves the efficiency of the parallel algorithm,and as the scale of the problem increases,the efficiency of the parallel algorithm increases.The research of parallel algorithm in this paper improves the forward speed and accuracy of RMT,and lays a foundation for the realization of subsequent fast inversion algorithm.2)The forward and inverse problems of RMT can be attributed to the forward and inverse problems of the Helmholtz equation of electric field or magnetic field derived from Maxwell field equations.The physicals information neural network(PINN)can solve these problems.Taking the one-dimensional forward calculation of RMT as an example,the PINN is compared with the multi-layer fully connected neural network(FCNN),and the generalization ability of the two models is analyzed.Taking the one-dimensional inversion calculation of RMT as an example,the inversion performance of FCNN is tested.The results show that PINN does not need to generated sample data,and can deal with models such as continuous medium models that traditional neural networks cannot handle.When dealing with forward problems,PINN adds physical equations to network training,Improved the generalization ability of the algorithm.FCNN inversion can achieve good results even in the presence of noise,and has certain robustness.In general,it is feasible for PINN to apply neural network to RMT numerical simulation and inversion.