Research On Magnetotelluric Intelligent Inversion Algorithm Based On Deep Learning

Magnetotelluric(MT)sounding method is an important geophysical exploration method,which can infer underground conductivity information through the distribution of natural alternating electromagnetic fields.It has been widely applied in mineral oil and gas exploration,geodynamics,environmental engineering and other fields.Inversion plays an important role in geophysical data processing and is a bridge connecting geophysical observation data and interpretation.There exist nonlinear and ill-posed properties in the inversion problem of MT data.Most of the existing methods are based on iterative inversion of objective function optimization theory.This type of method depends on the choice of the initial model and is easy to fall into a local minimum.Probabilistic inversion methods,despite their great potential in uncertainty quantification,still remain a formidable computational task.Aiming at the above problems,focusing on the current rapid development of artificial intelligence algorithms,starting from the non-linear nature of MT inversion,an intelligent inversion algorithm of MT data based on deep learning is proposed in this paper.The main research works and results are as follows:For the 1D inversion of MT data,we propose a hybrid model(CNN-LSTM)that combines convolutional neural network(CNN)and long short-term memory network(LSTM)to establish the complex nonlinear mapping relationship between resistivity model and MT response data.CNN will be used to extract the shallow-deep features in the electromagnetic response data,and LSTM will further process the timing information from the CNN layer and reconstruct the resistivity model.In order to improve the generalization ability and inversion accuracy of the network,a dynamic random probability data set generation method is proposed to ensure the sufficient and completeness of the data set.ADAM optimization algorithm is used to optimize the network training process.Theoretical and measured data are used to test the effectiveness of CNN-LSTM.The results show that(1)the CNN-LSTM hybrid network has extremely high accuracy and fast speed,which can achieve accurate imaging of complex1 D resistivity models,and this method still shows good robustness to the inversion of noisy data.(2)Compared with the traditional deep learning inversion methods such as convolutional neural networks and fully connected deep neural networks,the CNN-LSTM hybrid network model has higher accuracy.(3)The CNN-LSTM hybrid network can realize fast 1D inversion of measured data,whose results are consistent with the inversion results of OCCAM,and are verified by drill core data.Aiming at the 2D inversion of MT data,we propose an "end-to-end" inversion method based on the fully convolutional networks(FCN),which builds the mapping from apparent resistivity and phase data(inputs)to resistivity model(output)directly.In order to make full use of the weight sharing mechanism of FCN,only simple resistivity models are used for training,which not only cuts back the modeling time,but also reduces the difficulty of network training.In the inversion prediction stage,the trained network is verified through the inversion of unknown resistivity anomalous bodies models.We use several theoretical experiments to verify the effectiveness of the FCN method,and the results show that(1)the FCN inversion method has good generalization ability and anti-noise performance,and can realize accurate positioning and imaging of unknown resistivity anomalous bodies models.(2)Compared with the single TE or TM mode FCN inversion,the joint inversion of TE and TM mode has higher accuracy.(3)Compared with the traditional least squares inversion method,the FCN inversion method can describe the boundary of the anomalies more clearly.Finally,the FCN inversion method is applied to the actual data,and the inversion results show the same characteristics as the OCCAM inversion results,which shows the practicability of this method.The above-mentioned research has realized 1D and 2D high-efficiency and highprecision inversion of MT through deep learning technology,which provides the possibility for real-time estimation of underground resistivity distribution.