Xuben Wang

The Wenchuan Ms 8.0 earthquake on May 12, 2008 (WCEQ), and the Lushan earthquakes (LSEQ) on April 20, 2013, and June 1, 2022 (Ms 7.0 and 6.1, respectively), occurred in the middle and southern segments of the Longmenshan (LMS). The asymmetric spatial distribution of aftershocks following these events resulted in the formation of a seismically quiescent zone—designated as the “seismic gap” (SG)—between their rupture areas. Magnetotelluric sounding is employed to map the three-dimensional electrical structure of the middle-southern LMS region to elucidate the geophysical mechanisms underlying SG formation. Two fragile zones (A and B) act as a buffer zone are identified along the fault system: Zone A restricting northeastward propagation of LSEQ aftershocks, while Zone B constrains southwestward expansion of WCEQ aftershocks, collectively shaping the current SG configuration. Zone A exhibits a shallow surface fracture structure combined with partial melts derived from crust and the upper mantle, forming a low-resistivity anomaly zone (C2). Significant spatial variations in principal electrical axis orientation, maximum compressive stress direction, fault kinematics, and crustal block movement are observed across the SG boundary. These differences indicate the presence of tensile stress components along SG-margin faults, promoting crustal tearing processes. Consequently, upper crustal materials within Zone B undergo fragmentation and fluid infiltration, generating additional low-resistivity anomalies (C1 + C3) associated with enhanced crustal hydration.

Identifying mafic–ultramafic intrusions is critical for exploring magmatic sulfide deposits, especially those with abrupt morphological changes that may host mineralization. A dynamic iterative reweighting matrix inversion method enhances the resolution of magnetic susceptibility anomaly reconstruction and improves the characterization of intrusion morphology. This method initially constructs a weighted matrix based on cross-correlation coefficients between observed data and theoretical responses generated by a unit source. It then optimizes a dual-objective function balancing data fitting and weight distribution, iteratively refining the weighted matrix to adapt to subsurface structures. Finally, the optimized weighting matrix is used as a constraint to solve for the magnetic susceptibility anomaly model. A synthetic model test demonstrates significant improvement in detecting a single magnetic body with varying dip angles at different depths. Applied to the Tianfang magmatic sulfide deposit in Sichuan Province, China, this method successfully maps intrusion distribution and morphology using unmanned aerial vehicle aeromagnetic data. Two potential favorable ore-prospecting areas are identified: (1) transitional zones between steep and gentle dips and (2) neck regions of bottle-shaped intrusions. The integration with polarizability and resistivity anomalies led to two high-confidence drill targets, with drilling confirming disseminated sulfide mineralization at predicted depths of 340–400 m. These results indicate the effectiveness of dynamic reweighting matrix inversion in guiding magmatic sulfide deposit exploration. The method may also be useful for exploring other mafic–ultramafic-hosted deposits and magnetite ores.

Conventional numerical methods for computing three‐dimensional (3D) airborne transient electromagnetic (ATEM) forward modeling suffer from low computational efficiency and high computational cost. Although deep learning techniques have achieved some progress in accelerating low‐dimensional ATEM forward modeling, high‐accuracy and efficient 3D forward modeling has yet to be realized. To address this gap, we propose a deep learning–based approach for rapid and accurate 3D ATEM forward modeling. To simulate complex and heterogeneous geological conditions, we construct a large‐scale multi‐structure data set that incorporates most common subsurface features, including folds and faults. To address the extremely high computational cost required by transformers when processing 3D data, as well as the limitation of convolutional networks in globally modeling geoelectric structures—given that 3D ATEM forward modeling requires consideration of the entire model—we develop a network architecture based on the receptance weighted key value (RWKV). By employing bidirectional quadratic expansion and bidirectional weighted key value (Bi‐WKV) operations, the network effectively captures strong geometric relationships and structural continuity within 3D distributions. Furthermore, we introduce a transceiver altitude modulation mechanism, enabling the network to accurately handle variations in forward modeling induced by changes in transceiver altitude. Comprehensive experiments demonstrate that the proposed method achieves fast and high‐precision 3D ATEM forward modeling, exhibits sensitivity to transceiver altitude, and shows applicability to realistic geoelectrical models.

The semi-airborne transient electromagnetic method (SATEM) is capable of conducting rapid surveys over large-scale and hard-to-reach areas. However, the acquired signals are often contaminated by complex noise, which can compromise the accuracy of subsequent inversion interpretations. Traditional denoising techniques primarily rely on parameter selection strategies, which are insufficient for processing field data in noisy environments. With the advent of deep learning, various neural networks have been employed for SATEM signal denoising. However, existing deep learning methods typically use single-mapping learning approaches that struggle to effectively separate signal from noise. These methods capture only partial information and lack interpretability. To overcome these limitations, we propose an interpretable decoupled representation learning framework, termed DREMnet, that disentangles data into content and context factors, enabling robust and interpretable denoising in complex conditions. To address the limitations of convolutional neural network (CNN) and transformer architectures, we utilize the receptance weighted key value (RWKV) architecture for data processing and introduce the contextual weighted key value (Co-WKV) mechanism, which allows unidirectional weighted key value (WKV) to perform bidirectional signal modeling. Our proposed Cover Embedding technique retains the strong local perception of convolutional networks through stacked embedding. Experimental results on test datasets demonstrate that the DREMnet method outperforms existing techniques, with processed field data that more accurately reflects the theoretical signal, offering improved identification of subsurface electrical structures.

Deep learning has improved automated seismic phase picking in recent years. However, many pickers are trained on a single dataset and often fail to generalize when deployed on out-of-distribution (OOD) data. Variations in earthquake magnitude, propagation distance, sensor instrumentation, and ambient noise between training and target domains lead to significant performance degradation under OOD conditions. To address this challenge, we propose a new framework to reduce performance degradation under OOD conditions based on a pipeline of seismic simulation -- phase labeling -- site-conditions simulation -- transfer learning. First, we use the seismic simulation tool AxiSEM3D to establish a 3-D waveform simulation, covering local, regional, and teleseismic scales. Next, we obtain precise P- and S-phase labels by computing theoretical arrival times from velocity models and refining these onsets with traditional automatic picking algorithms, ensuring high-fidelity phase annotations. Then, based on actual site conditions, we simulate instrument responses and synthesize ambient noise constrained by PPSD analysis. This gives us station-specific, noisy waveforms that closely match real observational conditions. Finally, we employ transfer learning to fine-tune a phase picker on this specific synthetic dataset, thereby enhancing the picker's performance in new conditions. The proposed framework aims to improve the ability of deep learning models to pick phases under OOD conditions. It enables reliable performance across regional variability and instrumentation differences without large-scale manual relabeling. It also reduces the amount of real data needed for training, making it useful for small datasets and leaving more data to analyse. Overall, this work introduces a transferable methodology for seismic phase picking under distribution shift and shows that physics-informed data augmentation combined with targeted transfer learning can effectively decrease OOD performance degradation, thereby increasing the applicability of deep-learning-based phase picking.

The transformation of the gravity field into the gravity gradient data is an important topic in geophysical research, playing a key role in mineral boundary delineation and geological structure interpretation. However, traditional wavenumber-domain and spatial-domain transformation methods are often susceptible to noise and boundary effects, thereby limiting the accuracy, stability, and adaptability of the transformed gravity gradient data in complex geological settings. In this study, we introduce a new approach to the solution of this problem based on multi-task learning U-Net deep Neural Network. First, we propose an innovative approach to construct geological density model datasets using probabilistic constraints. Next, we utilize the Physics-Guided Multi-Task (PGMT) Network, a multi-layer CNN module, to enable the transition from single-task to multi-task outputs, capturing interdependencies among gravity gradient components. Finally, we incorporate the Laplace equation into the network architecture as a physical constraint, ensuring the three gravity gradient components adhere to the solid physical relationships. Numerical studies show that this network effectively alleviates noise and boundary effects. The determination coefficient (R²) is above 0.98, significantly reducing the root mean square error (RMSE), thereby greatly improving the prediction accuracy and robustness. After verification using the total tensor gradient gravity data of the Nordkapp Basin in Norway, the average error between the predicted gravity gradient data and the on-site measurement results is less than about 1%. These results demonstrate that the PGMT Network provides high precision and reliability in transforming low-order gravity data into high-order gravity gradient data, confirming its effectiveness for gravity field transformations.

Phase transitions among Mg2SiO4 and its high-pressure polymorphs (wadsleyite and ringwoodite) are central to mantle dynamics and deep-mantle material cycling. However, the locations and Pressure-Temperature (P-T) dependences of these phase boundaries remain debated, largely due to experimental limitations at extreme conditions and the high computational cost of first-principles free-energy calculations. Here, a machine-learning-potential driven workflow combining non-equilibrium thermodynamic integration (NETI) and two-phase coexistence simulations is employed to enable large-scale, long-timescale molecular dynamics sampling. Within this workflow, the melting curve of forsterite is evaluated and a complete P-T phase diagram is constructed. Relative to conventional ab initio approaches, this strategy reduces computational expense while retaining thermodynamic consistency in phase-stability assessment. The workflow is applicable to efficient evaluation of phase stability and thermodynamic properties in deep-Earth silicate systems.

Magnetotellurics (MT) is a widely used geophysical exploration method that has proven effective in geological surveys, resource exploration, and studies of deep Earth structure and dynamics. However, MT time-series data are inherently susceptible to various types of noise due to the reliance on natural source fields. Although traditional denoising methods, such as wavelet transform-based techniques, have shown certain efficacy, they often suffer from limited robustness and complex processing workflows when dealing with random and heterogeneous noise.Deep learning approaches have demonstrated strong capabilities in temporal signal modeling and have achieved promising results in time-series denoising and prediction tasks. In this study, we propose a novel deep learning framework for MT time-series denoising, termed DMTnet. The core component of DMTnet incorporates the multi-head attention mechanism of the Swin Transformer, enabling dynamic feature weight optimization across sliding windows. This design allows for end-to- end denoising of MT time-series data while preserving critical signal characteristics.Experimental results on both synthetic and real-world datasets show that the proposed method achieves superior denoising performance under complex noise conditions. Compared with several state-of-the-art deep learning denoising models, DMTnet consistently outperforms them across multiple evaluation metrics. Furthermore, frequency-domain analyses, including apparent resistivity and phase responses, demonstrate the accuracy and reliability of the denoised results. These research results indicate that DMTnet can significantly improve the quality of MT data and provide a solid foundation for the subsequent inversion and interpretation processes.

A 3-D gravity inversion can reconstruct subsurface density distributions from gravity observation data and is widely used in geological structure research and resource exploration. However, traditional methods often face challenges such as strong nonuniqueness and low depth resolution when inverting 3-D models from 2-D observation data. Currently, the issue of nonuniqueness is excessively pronounced, causing significant difficulties in accurately resolving complex subsurface structures. To address this, this study proposes a novel joint inversion method of gravity and gradient tensor data that integrates adaptive Lp-norm with an improved depth-weighted strategy. By calculating the first-order derivative magnitudes of model parameters in the $x$ -, $y$ -, and $z$ -directions and substituting them into the $p$ function, dynamic optimization of the norm selection is achieved. Simultaneously, nonlinear least-squares are employed to fit the gravity and gravity gradient kernel functions using a depth-weighted function, thereby optimizing the parameters within the depth-weighted function. This method significantly enhances the inversion stability and the characterization accuracy of geological structure, providing an efficient and robust solution for 3-D inversion of complex geological bodies. Synthetic model tests demonstrate that this method can more accurately characterize the location, size, and density amplitude of anomalous bodies compared with traditional inversion methods. Further application of this method to actual data from the Vinton Salt Dome in Louisiana, USA, yields inversion results that are similar to existing studies, validating the effectiveness of the method.

Gravity inversion, as a critical technique in gravity exploration, is widely used for reconstructing three-dimensional (3-D) subsurface density structures in mineral resource exploration, deep structural analysis, and crust-mantle isostasy studies. Currently, deep learning methods for gravity inversion based on the U-shaped Convolutional Neural Network (U-Net) architecture often lack sufficient information regarding constraints in gravity, causing significant discrepancies between observed and recovered gravity data. Such inconsistencies directly affect the confidence interval of the inverted density model and its spatial distribution in the subsurface. To address these issues, we propose an inversion network architecture integrating explicit linear Gravity Kernel Matrix (GKM) layers for joint gravity and gradients, termed the Kolmogorov-Arnold theorem and Shifted-window Transformer-based network (KAST-Net). It consists of a two-dimensional (2-D) Kolmogorov-Arnold theorem-based encoder-decoder branch, a 3-D Shifted-window Transformer-based encoder-decoder branch and a series of the GKM layers between the 2-D branch and 3-D branch. Synthetic inversion model tests demonstrated that the proposed network achieves higher inversion accuracies for model boundary delineation and confidence intervals of inverted density, compared to the Residual U-shaped Convolutional Neural Network++ (ResU-Net++), the Kolmogorov-Arnold U-shaped Convolutional Neural Network (KANU-Net) and two other classical constrained inversion methods based on regularization. In addition, incorporating graph convolution into KAST-Net enables effective handling of non-uniform gravity data with missing values, maintaining superior performance even under data sparsity conditions. Finally, we applied the proposed KAST-Net to 3-D density imaging of the Vinton Salt Dome in southwestern Louisiana, USA. KAST-Net not only accurately estimates the depth of the caprock, located between 160 and 360 m, but also successfully identifies the underlying salt layer below 520 m, a new insight that outperforms other existing methods in both accuracy and reliability.