Conventional linear iterative methods for magnetotelluric sounding (MT) suffer from considerable limitations related to difficulties in selecting the initial model and local optima. On the other hand, conventional intelligent nonlinear methods exhibit slow convergence and low accuracy. In this study, we propose an inversion strategy based on the deep learning (DL) deep belief network (DBN) to realise the instantaneous inversion of MT data. A scaled momentum learning rate is introduced to improve the convergence performance of the restricted Boltzmann machine during the DBN pre-training stage, and a novel activation function (DSoft) is introduced to enhance the global optimisation capability during the DBN fine-tuning stage. To address the difficulty in designing the sample data when prior information is lacking, we employ the k-means++ algorithm to cluster the MT field data and use the clustering results as the prior information to guide the construction of the sample dataset. Then, based on the proposed DBN, we ensure end-to-end mapping directly from the apparent resistivity to the resistivity model. We implement two groups of experiments to demonstrate the validity of both improvements on the DBN. We consider six types of geoelectric model from the test set to demonstrate the feasibility and effectiveness of the proposed DBN method for MT 2D inversion, which we further compare with the well-known least-square regularisation method for several extended geoelectric models and field data. The qualitative and quantitative analyses show that the DL inversion method is promising as it can accurately delineate the subsurface structures and perform rapid inversion.
11 August 2023 in IEEE Transactions on Geoscience and Remote Sensing
Despite exhibiting excellent inversion results for synthetic data in Magnetotelluric (MT) inversion, applying deep learning (DL) to directly inverting MT field data remains challenging. In this study, different from most previous works that mainly focus on generating massive representative resistivity models to cover the solutions of the field data or constructing a strong network by employing advanced DL techniques, we provide a new perspective in that a multi-window Savitzky-Golay (MWSG) filter is proposed to first smooth the apparent resistivity and phase derived from the MT field measurements before network prediction. This smoothing operation aims to promote the actual apparent resistivity and phase to be close in morphology and smoothness to the training input data, i.e. to adapt the field data to the training sample data. Then, the smoothed apparent resistivity and phase, instead of the original ones, are fed into the well-trained network for instantaneous inversion. Because we create a set of layered resistivity models with gradual-changing resistivity to act as desired output during network training, it together with the proposed MWSG filter enables this work to achieve smooth inversion. Besides, we introduce Swin Transformer (SwinT) to improve the efficiency of MT DL inversion, based on which a physics-informed neural network (PISwinT) is implemented to enhance the generalization capability. We demonstrate the proposed PISwinT-MWSG smooth inversion method in both synthetic and field MT cases, and it is expected to improve the adaptability and practicability of the DL method to directly solve the inverse problems in MT surveys.
11 September 2024 in IEEE Transactions on Geoscience and Remote Sensing
An ongoing trend seeking to incorporate forward modeling, which involves the physical laws of wave propagation, into the network architecture to improve the generalization capability of the deep learning (DL) inversion method has showcased promising applications. However, directly embedding the time-consuming two-dimensional (2-D) magnetotelluric (MT) forward modeling solved by conventional numerical algorithms to facilitate physics-guided DL MT inversion, which usually necessitates millions of forward operations during a complete training session, is challenging. Hence, in this work, we develop a physics-guided DL inversion method (PGWNet) by constructing a W-shaped DL model and performing a multitask learning strategy. The DL model consists of one encoder and two decoders, where the two decoders are independent of each other and share the encoder. During the training process, two decoders are firstly optimized independently by minimizing the model misfit, quantifying the discrepancy between the predicted and labeled resistivity models, and the data misfit, quantifying the discrepancy between the predicted and labeled MT responses, respectively. When model and data misfits back-propagate to the encoder, they are combined to jointly optimize the encoder. Moreover, to ensure practical application effect, this work builds a set of random synthetic resistivity models with gradually varying resistivity values to delineate realistic subsurface structures. We substantiate the developed PGWNet inversion method using synthetic and actual MT data and benchmark it against a fully data-driven DL inversion method and the conventional least-squares gegularization inversion method. It is anticipated to promote the practicability and applicability of the DL inversion method in practical MT prospecting scenarios.