Geophysical Journal International

SummaryIn a growing number of studies, researchers have presented induced polarization as a dielectric-related phenomenon arising from the complex nature of the permittivity in the constitutive law associated with the displacement current. Using Ampères’s law, they then write the conductivity as a complex number with an imaginary part related to the real part of the permittivity mixing Maxwell-Wagner-Sillars polarization and induced polarization phenomena. This presentation may be misleading to students and new researchers in the field. In terms of underlying physics, induced polarization phenomena have nothing to do with the displacement current density and Maxwell-Wagner-Sillars polarization should not be considered as an induced polarization mechanism. At the microscopic level, low-frequency polarization arises because the ionic fluxes responsible for the local electrical current are coupled not only through the long-range Coulombic effect but also through their physical interactions associated with diffusion phenomena driven by the random thermal motion of charge carriers. The thermodynamic force associated with ionic migration is not the electrical field alone but electrochemical potential gradients (induction effect can be accounted for, if needed, in the electrical field component of the Nernst-Planck equation governing the fate of ions in the pore space and along the surface of the mineral grains in their so-called electrical double layer). This has been known since the seminal papers of D.J. Marshall & T.R Madden in the 50s and Vinegar & Waxman in the 80s. This point seems, however, forgotten in recent studies leading to some misconceptions preventing a mechanistic understanding of the induced polarization processes and broad-band spectroscopic data.

SummarySufficiently accurate 3-D seismic wave modelling in the presence of topography remains computationally demanding, particularly when curvilinear discretizations are applied throughout the full domain. We develop an overset Virieux-Lebedev finite-difference framework that confines the curvilinear Lebedev-grid (LG) discretization to a near-surface patch and couples it to a standard staggered-grid (SSG) scheme in the deeper region through wavefield exchange across an overlap zone. The approach aims to reproduce traction-free free-surface effects and topographic scattering while reducing the cost associated with domain-wide curvilinear LG finite-difference operators. We first verify the underlying mimetic finite-difference implementation using homogeneous half-space benchmarks with both flat and Gaussian topographic free surfaces by comparison with reference solutions. We then assess the overset coupling using a series of verification tests, including a homogeneous Cartesian benchmark,a layered model with topography and impedance-contrast interfaces, and a modified GO_3D_OBS crustal model. Across these experiments, the receiver seismograms and wavefield snapshots exhibit no discernible artefacts attributable to the overset coupling interface, and the time-frequency envelope and phase misfit measures indicate close waveform agreement. Runtime profiling further indicates substantial efficiency gains when the LG overset thickness is held fixed, and a representative GO_3D_OBS experiment achieves a marked reduction in wall-clock time relative to a full-domain LG simulation. These results demonstrate that the proposed overset strategy enables efficient and sufficiently accurate 3-D wave propagation modelling for complex topographic settings relevant to waveform-based imaging and inversion.

SummaryKnowing and monitoring the spatial distribution of ambient seismic noise sources is essential for correlation-based investigations in seismology. The omnipresent primary and secondary microseismic noise characterizes the ocean-generated noise of the seismic spectrum and has been studied through observations and modeling to better constrain the source mechanisms, the source regions, and their temporal variability. In this study, we demonstrate the potential of a single six-degree-of-freedom (6-DoF) station, observing co-located translational and rotational ground motions, to determine and continuously monitor the source directions of the secondary microseismic wavefield. We harness direct rotational ground motion data obtained with the ROMY ring laser array. The effective array-like capability of a single 6-DoF station to separate Love and Rayleigh waves, both constituents of the secondary microseismic wavefield, by their polarization enables the independent estimation of the dominant source directions and the monitoring of their seasonal evolution. An anticipated seasonality is evident, with a dominant source direction for Love (254○) and Rayleigh (277○) waves in winter, whereas in summer no clear dominant direction emerges due to an absence of strong nearby sources. We compare 6-DoF-based backazimuth estimates with those of seismic array beamforming and find close agreement concerning the dominant source directions, particularly in the case of migrating pelagic storms in the North Atlantic associated with strong ocean wave activity over several days. In winter, a systematic bias of about 25○ to 35○ for Rayleigh waves and 50○ for Love waves is identified between the 6-DoF-based backazimuth estimates and those inferred for significant wave heights obtained from satellite altimetry and model-based seismic noise source maps. Our results provide a proof-of-concept for the direction monitoring of the secondary microseismic noise for both Love and Rayleigh waves using direct rotational observations, thereby reducing reliance on array-derived estimates. The advent of portable high-sensitive rotation sensors will facilitate constraining spatial seismic noise source distributions and temporal variations therein with future 6-DoF station networks potentially replacing or complementing traditional seismic array deployments.

SummaryElastic full waveform inversion (EFWI) provides high-resolution subsurface P- and S-wave velocity models essential for formation lithology and fluid characterization. However, its strong nonlinearity makes it highly susceptible to the local-minima entrapment when low-frequency seismic data or an accurate initial model is unavailable. Furthermore, the coupled sensitivity of multi-component seismic data to distinct elastic parameters introduces a multi-parameter crosstalk effect that further degrades inversion accuracy. Although well-log data contain rich, inherently decoupled prior information of subsurface elastic parameters, a fundamental dimensional mismatch between 1D well logs and the 2D/3D inversion domain prevents their direct integration into conventional EFWI. To overcome these limitations, a joint physical-model- and multi-source-data-driven (JPM-MSDD) EFWI framework is proposed. The correlation between common mid-point (CMP) gathers and 1D vertical velocity profiles is exploited to resolve the dimensional mismatch, enabling well-log data to directly serve as training labels. A neural network is trained to map consecutive adjacent CMP gathers of multi-component seismic data to 1D P- and S-wave velocity profiles under a semi-supervised learning scheme, in which well-side CMP–log pairs form the labeled dataset while model-driven EFWI furnishes physics-consistent pseudo-labels for unlabeled gathers. In this way, the proposed method is able to reduce the reliance on large labeled datasets and enhance the physical interpretation of inversion results. To further reduce the nonlinearity of EFWI, the spatially related information of seismic data is incorporated through position embedding on seismic traces. By integrating multi-source data, including well-log data, multi-component seismic data, and spatial information of seismic traces, into EFWI process, even in the absence of low-frequency components of seismic data, a very good low-wavenumber model can be inverted for conventional model-driven EFWI to prevent the inversion from falling into local minima. Additionally, the proposed method can mitigate the crosstalk of multi-parameters to a certain extent. The synthetic and field data examples demonstrate that the proposed joint physical-model- and multi-source-data-driven EFWI can invert vp and vs models effectively.

SummaryHigh-frequency (> 1Hz) coda waves consist of multiply scattered seismic energy that follows direct body wave arrivals and exhibit high sensitivity to small-scale lithospheric heterogeneities and attenuation. The decay and shape of coda wave envelopes are mainly influenced by intrinsic attenuation, associated with irreversible energy loss due to anelastic processes, and scattering attenuation, resulting from wave redirection caused by heterogeneity. However, these two effects are strongly coupled in observations, making their separation challenging. Here, we apply a Conditional Variational Auto-Encoder (CVAE) to 28,393 high-frequency, vertical-component coda wave envelopes recorded across the eastern margin of the Tibetan Plateau. By conditioning the model on epicentral distance, the trained CVAE extracts two latent variables that characterize the dominant envelope-shape variations independent of the first-order distance effect. The first latent variable mainly modulates P/S amplitude ratio and post-peak decay behavior, showing trends qualitatively consistent with intrinsic attenuation effects. Its spatial distribution reveals lower values within the stable Sichuan Basin and high values across the tectonically active Tibetan Plateau side, consistent with previous attenuation studies. The second latent variable primarily controls envelope broadening and peak delay, consistent with scattering-related energy redistribution. Its spatial pattern suggests enhanced scattering-related broadening in the region affected by loose sedimentary cover in the basin and fault-related fracturing in the plateau. The synthetic data test further supports the qualitative interpretation of the latent variables, while also showing cross-sensitivities between them. These results indicate that the CVAE can extract attenuation-related envelope-shape features, rather than directly decoupling intrinsic and scattering attenuation parameters. Our findings demonstrate the potential of CVAE-based approaches for extracting physically interpretable attenuation-related patterns from large coda envelope datasets in tectonically complex regions.

SummaryAccurate three-dimensional (3D) density models aid dynamics and oil exploration, yet inversion in complex regions faces non-uniqueness and noise. Due to the non-uniqueness and noise sensitivity of the solution process, accurately resolving spatial density variations from gravity data remains a major challenge. Wavelet decomposition is applied to the separation of gravity signals, and inverting the processed signals can effectively reduce non-uniqueness. However, wavelet multiscale decomposition can suffer from inter-layer crosstalk and boundary blurring due to incomplete signal separation. To address this, we proposed the Layered-Integrated Density Inversion Method (LID). The process begins with a layered inversion using wavelet decomposition. This integrated model serves as the initial model in the subsequent simulated annealing inversion. Simulation experiments showed that the Root Mean Square Error (RMSE) of the LID results is 0.056 g/cm3, representing a 63.47% improvement over the traditional method, which is based on wavelet layered inversion. In noise resistance experiments, the RMSE of the LID results was 0.077 g/cm3, a 54.67% improvement over the traditional method. Applying the method to the South China Sea (SCS) using Surface Water and Ocean Topography (SWOT) satellite vertical gravity gradient data, we identified three prominent geological features: (1) traces of seafloor spreading along the mid-ocean ridge of the SCS; (2) fractures generated during subduction of SCS plate; (3) anomalous signals detected southeast of Palawan, which possibly correspond to remnants of the Paleo-Pacific plate. These findings align with existing theoretical models, providing independent evidence for rift dynamics and post-spreading crustal evolution of this passive continental margin.

SummarySeismic moment tensors provide simple representations of a wide range of event types, including earthquakes, volcanic events, landslides, and explosions. Estimating the six parameters of a seismic moment tensor is subject to inaccuracies in the assumed Earth model, whether it is characterized by a simple layered structure (1D) or by heterogeneities (3D). Using recorded data and seismic wavefield simulations in 3D Earth models, we estimate double-couple and full moment tensors using MTUQ software, and then we quantify the differences among moment tensors due to the different Earth models. We establish new capabilities for implementing 3D Earth models from the EarthScope Earth Model Collaboration (EMC) into the wave propagation code SPECFEM3D Globe. The three workflow components—EMC, SPECFEM3D Globe, and MTUQ—are publicly available, thereby maximizing possibilities for access, reproducibility, and future enhancements. We demonstrate the source estimation workflow in the region of Alaska, using 5earthquakes and 53D Earth models. Using misfit measures from the moment tensor estimation, we can quantify the performance of both the moment tensor and also the underlying Earth model. For this limited data set of 5 well-recorded events, we demonstrate how to examine basic questions of source variability, tomographic model evaluation, and the reliability of non-double-couple components of moment tensors. The procedures enable quantification and visualization of moment tensor uncertainties due to Earth models, while also offering direct comparison of tomographic models with an independent reference set of data.

SummarySeismic anisotropy observations can constrain flow and deformation in the lowermost mantle (D″). D″ deformation is often attributed either to changes of mantle flow from mostly horizontal to upwelling near deep mantle plumes or edges of the two antipodal large low-velocity provinces (LLVPs), or to strong deformation in slab-dominated deep mantle regions. A unified understanding of deep mantle flow and its drivers, however, is still developing. We investigate D″ anisotropy beneath Australia using a novel approach combining array processing and shear-wave splitting measurements applied to core-traversing seismic waves which reflect off the underside of the core-mantle boundary up to two times (called SKS, S2KS, and S3KS). Strong differences in splitting between pairs of phases that sample the upper mantle in a similar way but sample different portions of the D″ layer (e.g., SKS-SKKS or SKS-S3KS) can be interpreted as due to a contribution from lowermost mantle anisotropy to the splitting of one or both phases. Using this approach, we detect strong anisotropy in two locations at the southwestern edge of the Pacific LLVP, which we attribute to a change from mostly horizontal to upwelling flow. One of these locations coincides with a previously suggested mantle plume that has not yet reached the surface. We also identify anisotropy south of Australia, where seismic velocities are faster than average and slab remnants may be present. Beneath much of the Australian continent itself, we see little or no evidence for splitting discrepancies, which may be evidence for isotropic, or only weakly anisotropic, lowermost mantle. Our study region thus uniquely showcases seismic anisotropy in diverse deep mantle environments, associated with a possible deep mantle plume, the edge of the Pacific LLVP, and potential remnant slab material.

SummaryPhase and group velocities along specific ray directions are needed for qP, qSV and SH wave traveltime computation in tilted transversely isotropic (TTI) media. The phase and group velocities are not unique functions of the ray direction, but of the slowness direction, so efficient computation of the slowness direction from a given ray direction becomes necessary. The eigenvalue method and the generalized method have been proposed to facilitate the computation by formulating governing equations for the phase angle, which requires efficient root-finding algorithms. In this work, we address this problem in two-dimensional (2D) TTI media by applying Newton’s method. For the eigenvalue method, a set of two nonlinear equations for the qP and qSV waves and one nonlinear equation for the SH wave must be solved. For the generalized method, only one nonlinear equation of the phase angle needs to be solved for the qP, qSV and SH waves. Numerical experiments, including phase and group velocity computation and their application in first-arrival traveltime calculation, are performed to verify the effectiveness of the proposed methods.

SummaryThis study evaluates the performance of the velocity model adopted for the CyberShake 24.8 (CS24.8) study when used to constrain wave propagation in three-dimensional regional-scale physics-based simulations for seismic hazard estimates. The CS24.8 study was developed to estimate seismic hazard in a subdomain surrounding the San Francisco Bay Area (SFBA) in California, adopting a physics-based finite-difference scheme for frequencies up to 1 Hz; above that, a stochastic scheme with site-specific adjustments is used. The velocity model adopted in the CS.24.8 study was a modified version of the USGS regional velocity model developed for the SFBA. The evaluation of the velocity model is based on comparisons between simulated and recorded ground motions for 18 small-to-moderate local earthquakes. Our analysis focuses on two frequency ranges: 0-1 Hz for estimating wave-propagation uncertainties for users of the CS24.8 study, and 1-5 Hz to provide insight into the performance of the velocity model for future Cybershake studies in this region. Metrics based on Fourier amplitude spectra (FAS) and waveforms’ duration are used for quantitative evaluation of the velocity model. Two aspects of the simulated ground motions are analyzed: (i) the median and variability of the ground motion in the region and (ii) wave propagation effects for specific source-site pairs. For (i), the velocity model leads to an underprediction of the FAS ranging from 0.1 LN-units at 0.3 Hz to 0.5 LN-units at 1 Hz for the horizontal component, and an underprediction of the duration by a factor of 2. The underprediction can be explained by the 400 m/s minimum shear-wave velocity adopted in the CS24.8 velocity model, which is larger than the actual values in the soft marine quaternary sediments of the SFBA, where most stations are located. The spatial variability of the FAS from the simulations over the region is lower than that from the observations over the frequency range 0.3 to 1 Hz, suggesting that the 3-D velocity structure is too smooth. When extending the analysis up to 5 Hz, the underprediction pattern increases up to 0.7 LN-units, and the spatial variability of the ground motions increases, reconciling the gap observed at lower frequencies. For (ii), the evaluation shows that the 3-D simulations improve the accuracy of wave propagation effects for the FAS compared to the ergodic ground-motion models (GMMs) for frequencies less than 0.7 Hz and have similar accuracy up to 1 Hz, being the maximum frequency solved in the physics-based scheme of the CS24.8 study. When extending the analysis above 1 Hz, misrepresentations in the 3-D velocity model introduce noise into the simulated ground motions, leading to a less accurate estimate of the FAS at these frequencies compared to GMMs. Our results inform users of the CS24.8 study that the physics-based simulation (up to 1 Hz) offers performance comparable to or better than standard GMMs, while accounting for wave-propagation uncertainties. These findings can guide future refinements of the velocity model.

SummaryGeophysical electrical methods are increasingly being used in civil engineering to characterize and monitor cementitious materials. However, there is currently a lack of understanding of the role of their electrical surface conductivity and there is no quantitative model explaining their complex conductivity (induced polarization) spectra. Therefore, our goal is to propose and to validate a mechanistic model. We prepared 20 cement paste samples of well-established cement compositions (named CEMI and CEMV in the cement nomenclature) and 16 corresponding mortar samples (labeled MORI and MORV), all cured for 60 days, with water-to-cement (w/c) ratios ranging from 0.35 to 0.60. Complex conductivity spectra were measured at 21°C in the frequency range 10 mHz-45 kHz. For the cement pastes, both the in-phase conductivity and the magnitude of the quadrature conductivity increase systematically with the increase of the w/c ratio. The electrical properties of the mortars scale proportionally with those of the corresponding cement pastes, and the proportionality coefficient can be predicted from the volume fraction of cement and the model. We observe that the normalized chargeability is proportional to the quadrature conductivity, consistent with theoretical expectations. The relationship between the normalized chargeability and the surface conductivity and between the normalized chargeability and the Cation Exchange Capacity (CEC) are explained using a dynamic Stern layer model associated with the polarization of the inner component of the double layer coating the surface of the minerals. In other words, the dynamic Stern layer initially developed for colloidal solutions and geomaterials can be applied to cementitious materials opening new doors in their non-intrusive monitoring. To our knowledge, this is the first study to provide a physically-based interpretation of the complex conductivity spectra of cement pastes and mortars. These results demonstrate that induced polarization displays strong potentials for imaging water content and the Cation Exchange Capacity (CEC) (alternatively the specific surface area) of cementitious materials at various scales. This opens new perspectives regarding the quantitative non-invasive geophysical monitoring of cement and concrete for both civil and nuclear engineering applications.

SummaryUniform density-contrast assumptions in gravity-derived bathymetry produce substantial systematic errors. This problem stands out in regions with strong lateral density variation, such as the central–northern South China Sea. Conventional constant or simple vertically varying density models fail to capture these complexities. To overcome this limitation, a spatially varying density-contrast field is constructed by integrating multi-source geophysical data (crustal, gravity, and bathymetric data) using a back-propagation (BP) neural network. This field is incorporated into an adaptive Parker–based inversion, yielding a high-resolution bathymetric grid with Root Mean Square (RMS) improvements of 2.6 m over the constant-density approach, together with the smallest systematic bias. The most significant gains occur in shallow reef-dominated waters (0 to −1500 m), where relative RMS reductions reach approximately 9%. By coupling neural network-derived density modelling with physically rigorous inversion, the approach overcomes limitations of uniform-density assumptions while retaining interpretability, providing an efficient and reliable approach to high-precision seafloor mapping in geologically complex regions.

SummaryInstrument correction is a fundamental step in broadband seismology, yet it is commonly treated as a purely numerical operation stabilized by heuristic procedures. In practice, inverse filtering of instrumental responses is constrained by causality, stability, discretization, and uncertainty in instrument parameters, which jointly limit the recoverable frequency content and the physical interpretability of corrected ground motion. Here I present a unified framework for causal and uncertainty-aware instrument correction that explicitly formulates deconvolution as a constrained inverse problem in the digital domain. The proposed approach enforces causal realizability and bounded inverse behaviour while introducing regularization as a physically interpretable control of the bias–noise trade-off. Discretization effects arising from the mapping between continuous- and discrete-time responses are quantified and shown to induce systematic, frequency-dependent amplitude bias that interacts nonlinearly with regularization. I further extend the formulation to incorporate parametric uncertainty in the instrument response, propagating it through the inverse filter to derive confidence bounds on effective amplitude response and noise amplification. A set of diagnostic metrics is introduced to jointly characterize amplitude bias, noise amplification, effective bandwidth, and robustness under uncertainty. These diagnostics are combined into a data-driven decision framework that supports objective selection of the inverse filter and explicitly defines the frequency range over which instrument correction is reliable. Time-domain kernel diagnostics complement the frequency-domain analysis by ensuring causal behaviour and controlled temporal support. The framework is summarized in an end-to-end algorithmic workflow designed for reproducible application to broadband seismic data without reliance on ad hoc stabilization choices. By making physical constraints, trade-offs, and uncertainties explicit, the proposed methodology enhances the robustness and interpretability of instrument-corrected seismic waveforms and provides a principled foundation for downstream analyses in source studies, spectral characterization, and waveform-based investigations.

SummaryAirborne transient electromagnetic (ATEM) inversion relies on efficient forward modeling, yet conventional numerical methods struggle to balance computational efficiency and accuracy when handling undulating terrain and large survey datasets. We present a fast forward modeling approach using multi-scale dilated residual networks that takes two-dimensional conductivity profiles with embedded terrain information as image inputs and directly predicts electromagnetic response tensors across multiple flight altitudes. The network architecture employs progressively increasing dilation rates to capture multi-scale geological features while preserving spatial resolution. A dual-domain loss function combining log-normalized and linear domains balances the fitting weights across electromagnetic responses spanning seven orders of magnitude. We trained and tested the model on 100,000 synthetic samples of random terrain and geoelectric structures generated with the SimPEG three-dimensional finite volume method. The network achieves mean absolute percentage errors (MAPE) of 2.24%-3.57% across four terrain types (flat, slope, peak, and valley). Spline interpolation enables accurate prediction at arbitrary flight altitudes not included in training, with errors of 1.69%-3.53%. On a 10 km survey profile, the method achieves approximately 2000-fold speedup compared to SimPEG. This approach provides a practical forward modeling solution for rapid ATEM inversion over complex terrain, and the proposed multi-scale feature extraction strategy and dual-domain loss function design offer transferable insights for other geophysical machine learning applications.

SummaryTo address the challenge in seismic exploration where strong-energy ground roll severely interferes with effective signals and conventional suppression methods tend to damage these signals, this paper proposes a self-supervised Swin-Unet network method for ground roll suppression based on Fourier Positional Encoding and a bespoke masking strategy. This method operates without the need for clean label data. By employing a specially designed fan-shaped masking strategy, it disrupts the spatio-temporal coherence of the ground roll, thereby guiding the network to learn the intrinsic characteristics of the effective signals and reconstruct the data. The core innovation lies in the introduction of Fourier Positional Encoding, which overcomes the inherent low-frequency bias of the Transformer architecture. This significantly enhances the network’s capability to model and recover high-frequency effective signals. Experimental results on synthetic data and field 2D/3D seismic datasets demonstrate that the proposed method not only effectively suppresses strong ground roll but also surpasses the traditional f – k filtering method in terms of signal fidelity, particularly in preserving deep, weak reflections and high-frequency components. This showcases its robustness and potential for application in complex seismic data processing.

SummaryAn innovative method is presented for full 3D pressure and temperature dominated stress evaluations in the subsurface using tetrahedra-based analytical elements combined with the inflation point source solution. A mesh-free approach suitable to industry standard 3D flow simulation models (based on hexahedral cells) is obtained by representing the grid cells in tetrahedral elements, effectively preserving the 3D geometrical complexities of the reservoir. Contributions from neighboring grid cells are added via a Tartan grid representation of point sources that enables the spatial resolution to increase and the stress evaluations to be carried out in parallel, both of which significantly improve computational efficiency. The novel approach is demonstrated for synthetic low enthalpy geothermal models with clastic reservoir characteristics and varying degrees of geometrical and structural complexity. The method is shown able to accurately capture the effects of stress arching on complex faults causing reservoir throw and flow compartmentalization, and along the rim of the cold-water volume. Results of a synthetic geothermal development model of the heavily faulted Gullfaks field show the novel method to provide an accurate and computationally highly efficient approach for evaluating pressure and temperature dominated stress changes in structurally complex sedimentary reservoirs.

SummaryPore-water pressure and clay content have influence on porosity and bonding/cementing the grain boundaries, thus affecting elastic properties, strength, and other physical properties of rocks. Similar processes take place in cementitious materials, where in particular hydration plays crucial role. This process leads to changes in pore water composition, specific surface area (or alternatively Cation Exchange Capacity, CEC), and water content. Analytical expressions can be obtained between both the CEC and porosity and a hydration state variable. The hydration state variable can be in turn related to the hydration time. These phenomena can be assessed by geoelectrical methods, which have been used for a long time to observe the evolution of the textural properties and rheology of rocks, as well as of cementitious materials. However, such use has been so far rather qualitative. The aim of this study is to better describe the evolution of complex conductivity spectra (induced polarization) in relation to the hydration, using the recently developed dynamic Stern layer model to the hydration time through Powers model. Comparisons between the model predictions and literature data are used to test the suitability of the proposed solution. The model is verified using monitoring experiments of the complex conductivity spectra of several cements to study the evolution of both the in-phase and quadrature conductivity versus the hydration time. Although the current method is still semi-empirical, it makes it possible to analyze and understand the evolution of complex conductivity spectra of geomaterials and technogenic cementitious materials with a wide variety of geophysical applications in civil engineering, especially for dams and underground civil constructions.

SummaryWe have developed a physics-guided deep learning framework for geophysical inversion that incorporates Markov chain Monte Carlo (MCMC) sampling to assess the uncertainty associated with model parameters of interest. To enhance computational efficiency, a statistical sampling method is utilized to reduce the number of samples required while ensuring the training data remain both diverse and informative. As the inversion progresses iteratively, the training dataset is dynamically expanded using outputs from the stochastic sampler along with their corresponding forward responses. A supervised deep learning model is utilized, in which the Jensen-Shannon divergence is adopted as the loss function, and a Gaussian assumption is applied for analytical computation. We test the workflow on a seismic velocity model inversion, and successfully capture the geological features and velocity distributions, with results that closely match the reference model. Compared to the MCMC sampler applied to the whole data cube, the proposed workflow is more computationally efficient, as a small fraction of data is chosen using the active learning paradigm. This workflow is strongly generalizable and effective, making it suitable for a wide range of other inversion applications as well.

SummaryThe origin of seismic discontinuities in the Earth’s mid-mantle (∼700–1400 km) remains debated, with competing hypotheses attributing them to either partial melting due to water transport across the transition zone or compositional heterogeneities (subducted crust). Distinguishing between these scenarios has been hindered by the inability of standard imaging techniques to extract robustly the polarity of weak seismic reflections amidst noise and reverberations that contaminate mid-mantle reflections. Here, we introduce a novel signal processing framework that combines curvelet-based wavefield separation with extended multitaper deconvolution to resolve this polarity ambiguity. We validate this approach by applying it to a high-quality dataset of SS and PP precursors beneath the Central Pacific. This application yields the robust detection of a discontinuity at approximately 800 km depth, characterized by a sharp positive shear velocity contrast (δVS ≈ +4 − 5%) and a negligible density contrast. The observed positive polarity precludes partial melt or thermal plumes as primary causal mechanisms. Instead, the high-velocity, neutral-density signature is consistent with a layer of stagnant, subducted oceanic crust in thermal equilibration with the ambient mantle. These results demonstrate the efficacy of the deconvolution framework and provide direct seismic evidence for compositional stratification in the mid-mantle, supporting geodynamic models where viscosity increases facilitate the long-term preservation of recycled lithosphere.

SummaryGeodetic observations, such as GNSS and InSAR, are increasingly used to investigate co-seismic surface deformation. Efficiently and simultaneously resolving fault geometry and slip distribution from surface displacements is essential for understanding earthquake processes, accurately estimating seismic magnitude and comprehensively assessing seismic hazard. Current mainstream approaches typically rely on Bayesian inference, such as Monte Carlo sampling. However, these methods typically suffer from long burn-in periods, low computational efficiency, strong sensitivity to initial parameter values and step sizes. Given these limitations, conventional approaches may yield suboptimal fits for the observations. To overcome these limitations, we propose and develop a novel Tree-structured Bayesian Optimization method (TBO), integrated with Helmert Variance Component Estimation (HVCE), to jointly determine fault geometry and slip distribution. To rigorously assess the feasibility and reliability of the proposed approach, we test it using both synthetic and real earthquake data. In the synthetic tests, we evaluate its robustness under a variety of conditions, including different fault types, varying types and densities of geodetic observations, diverse sub-fault sizes and asperities, and complex multi-fault scenarios. Four sets of synthetic experiments are designed, and the results conclusively demonstrate that the proposed method achieves stable and reliable performance in inverting fault geometry and slip distribution. Furthermore, comparative analysis with existing methodologies shows that our approach yields substantially improved computational efficiency, significantly reduced sensitivity to initial conditions, and smaller misfits to observations. Finally, we apply the method to the 2021 Mw 6.4 Yangbi earthquake in Yunnan, China. The retrieved fault geometry and slip distribution successfully explain the fault kinematics and the observed surface deformation field, thereby confirming the applicability and robustness of the method for real earthquake event analysis.