SummaryThe application of machine learning to planetary geophysics is frequently bottlenecked by a scarcity of labeled data and the complexity of non-stationary seismo-acoustic signals. While deep learning offers powerful feature extraction, it requires massive datasets that are rarely available for phenomena like Martian quakes or deep-ocean volcanism. Consequently, researchers often rely on unsupervised machine learning or hand-engineered features that are typically impoverished or hard to interpret. To resolve this, we introduce Spec2VEC, a framework that transforms parameter-dependent TFRs into a robust, interpretable, and high-dimensional vector space. Spec2VEC treats spectrograms not merely as energy distributions but as structured, textured images, and extracts physics-informed, dataset-agnostic features without requiring massive training loops. The framework systematically extracts a feature vector using two novel pathways: (1) a Textural path that utilizes Hilbert space-filling curves to map 2D spectral patterns into 1D sequences for advanced information-theoretic analysis (e.g., permutation entropy, fractal dimension); and (2) a Spatial path driven by PDFSI (Probabilistic Distribution Fitting for Signal Identification), a novel probabilistic pre-conditioning engine that quantifies spatial connectivity by dynamically separating coherent signal modes from stochastic noise. We validate this feature space on a comprehensive benchmark of complex synthetic wavefields and two real-world examples: an Exotic Seismic Events Catalog (ESEC) and a Distributed Acoustic Sensing (DAS) dataset. Spec2VEC provides objective features for information discrimination and enables highly accurate classification of the diverse geophysical signal classes present in the ESEC catalog. Furthermore, application on the high data volume DAS dataset confirms its computational efficiency and domain-agnostic capability to isolate anomalous seismic arrivals from background noise. Ultimately, Spec2VEC bridges the interpretability of classical signal processing with the discriminative power of modern representation learning, providing a reliable, interpretable foundation for automated discovery across the geosciences.
SummaryAccurate three-dimensional (3D) seismic velocity modeling through seismic travel-time tomography using both active- and passive-source data provides critical underpinning models for seismicity monitoring and hazard assessment. Because travel-time tomography is an inherently ill-posed inverse problem, uncertainty quantification (UQ) of the estimated models using Bayesian methods is also important for reliable downstream interpretations and analyses. However, Bayesian inference for 3D tomography based on conventional grid-based representations faces the “curse of dimensionality” and severe computational bottlenecks. Consequently, rigorous Bayesian UQ for margin-wide 3D travel-time tomography has remained largely unexplored, despite its importance for reliable interpretation of large-scale velocity models. In this study, we propose a meshless 3D Bayesian travel-time tomography method that combines physics-informed neural networks (PINNs) with a neural representation of the velocity structure, enabling tractable and data-efficient Bayesian inference through function-space Particle-based Variational Inference (fParVI). To efficiently integrate passive-source data into the Bayesian estimation of the velocity structure, we introduce an analytical marginalization strategy that treats uncertain parameters related to the origin time and location as nuisance parameters, with passive-source relocation carried out in post-processing. We validated the capability of our approach for 3D problems through synthetic experiments. These experiments demonstrate its robustness against biased prior source locations compared to standard methods that suffer from overfitting. Furthermore, we applied the method to a real-world dataset off the Kii Peninsula, Nankai Trough, using over 380,000 travel-time picks from marine active-source surveys and natural earthquakes. Our probabilistic 3D ensemble successfully resolves key geological features, such as the Kumano Pluton, and provides data-consistent uncertainty maps that offer new quantitative insights into the reliability of previously debated deep low-velocity zones. The posterior mean hypocenters shifted mainly in the vertical direction by 10-15 km, consistent with the deterministic relocation results of a previous study. Finally, the continuous neural representation drastically reduces storage requirements for the entire ensemble velocity model, highlighting the scalability and data efficiency of the proposed framework for Bayesian seismic travel-time tomography.
SummaryA rigid connection between the optical fiber and the rock allows to properly measure ‘rock strain’ with Distributed Acoustic Sensing (DAS). We demonstrate this by running four interrogator units (IU) on a testbed with single-fiber patch cables being cemented into a groove in the concrete floor of Black Forest Observatory (BFO). The recorded signals are compared with the recordings of a calibrated Invar wire strain meter array that has been continuously in operation for the last decades at a distance of about 350 m to the testbed. In this way we compare the ‘fiber strain’ against ‘rock strain’ at frequencies below 0.2 Hz. We investigate this on the background of a previous study with DAS cables unreeled on the floor and loaded down by sand and sandbags, where we found a significant loss of strain amplitude between rock and fiber. Waveform similarity for strong earthquake signals is high with typical values of the normalized correlation coefficient greater than 0.95. The strain amplitude measured by all four IUs is very consistent and by about 10 per cent larger than the strain measured by the strainmeters. We attribute this minor discrepancy to spatially varying strain-strain coupling caused by the local topography. At frequencies up to 14 Hz we make an intercomparison of IUs, showing no significant variation with frequency. The scatter of strain amplitude in between channels which are spatially near to each other in the same fiber route is about ±10 per cent in most cases. The variation of median values in between different IUs and earthquakes is less than 5 per cent. By subtracting the common mode noise, which is coherent along the fiber route, we lower the background signal level to an rms-amplitude of 100 pstrain at 0.1 Hz and 5 pstrain at 1 Hz in a bandwidth of 1/6 decade for the best cases. This allows the detection of the marine microseisms during times of moderate amplitude level.
SummaryThe North Anatolian Fault (NAF) in the Sea of Marmara is characterized by varying mechanical behaviors and fluid manifestations along its path. In the present study, we investigate an earthquake sequence including a Mw 4.6 earthquake that occurred on November 27, 2013, in the partially creeping section of the NAF in the Sea of Marmara. Using onshore and offshore seismic data from October 2013 to August 2014, we document a sequence of highly-similar events prior to this earthquake, occurring within a zone where repeating earthquakes, likely indicative of creep, have been previously identified. This sequence coincides with the recently proposed existence of a several months-long slow-slip event (SSE) propagating westwards along the main fault trace at a rate of ∼0.5 km/day, based on the interpretation of onshore geodetic data and offshore surface sediment pore pressure data. The timings of the multiplet (i.e., the highly similar events), mainshock and decreases in pore pressure measured within the fault valley are compatible with the hypothesis that the SSE impacted first the multiplet-mainshock sequence and then the sub-seafloor hydrological network within the fault valley. Models of the SSE as a slipping patch propagating westward with a combination of right-lateral slip and tensile opening, show an increase in Coulomb stress of a few tenths of kPa at the position of the mainshock sequence. Despite limited calculated Coulomb stress transfer due to the SSE, our observations suggest that complex hydromechanical interactions exist between splay faults, such as those where the mainshock occurred, and the main NAF, and underline the interest of using multi-disciplinary data to study relationships between fluids and fault deformation.
SummaryWe investigate Reservoir-Triggered Seismicity (RTS) associated with the Castanhão Reservoir in northeastern Brazil, located within an intraplate region. Using continuous waveform data from a local seismic network (2009–2010), we construct a high-resolution earthquake catalogs of 227 events through template matching (PyMPA) and probabilistic location methods, with subsequent double-difference relocation improving spatial resolution for 103 events. The seismicity forms two clusters delineating a steeply dipping (∼80°) fault zone characterized as a reactivated strike-slip system with secondary transtensional/stepover structures within the Jaguaribe Shear Zone corridor, with moment magnitudes ranging from Mw 0.0 up to Mw 2.7. To assess triggering mechanisms, we model pore-pressure diffusion induced by reservoir impoundments. The estimated hydraulic diffusivity (∼0.080 m²/s) and spatiotemporal evolution of seismicity indicate a response governed by fluid pressure diffusion. Pore-pressure increases on the order of tens of kPa are sufficient to activate critically stressed fault segments, consistent with previous RTS observations. The seismic sequence exhibits swarm-like behavior, interpreted as progressive, fluid-driven rupture within a segmented fault network. We further analyze the temporal evolution of seismic strain using Seismic Strain Dynamics (SSD) and seismic regularity (dc) metrics, revealing patterns consistent with a multi-stage earthquake preparatory process. The largest event in the main cluster (Mw 2.1) occurs during a late-stage failure phase, following prolonged fault conditioning. Our results highlight the role of reactivated, critically stressed secondary strike-slip fault networks as hydraulically favorable conduits for pore-pressure diffusion and RTS. The study demonstrates that even moderate pore-pressure perturbations can induce delayed, swarm-type seismicity in intraplate settings, emphasizing the importance of integrating geological structure, hydromechanical processes, and seismic monitoring for improved hazard assessment near large reservoirs.
SummaryWe evaluated periodic pumping tests carried out in boreholes penetrating fractured gneiss in the research mine “Reiche Zeche”, Freiberg, Germany, employing a fractional diffusion equation, known as the generalized radial-flow (GRF) model. Water was injected into three borehole intervals, successively isolated by a double-packer probe. The three intervals correspond to distinct hydraulic conduits: (i) an induced axial fracture, (ii) a stimulated pre-existing radial fracture, and (iii) a fault zone intersecting the borehole at a near-normal angle. Pressure responses were monitored in sections of up to four monitoring boreholes. We evaluated the collected data by two spectral methods, injectivity analysis and interference analysis. The former exploits flow rate and pressure of the injection interval, whereas the latter relies on pressure records only, namely that of monitoring sections and of the injection intervals. In the GRF-framework, interference analysis yields a unique pair of flow dimension and hydraulic diffusivity, while the two observables of the spectral injectivity analysis, amplitude ratio and phase shift, only constrain admissible ranges of the three involved unknowns, flow dimension, hydraulic diffusivity, and conductivity. We addressed this under-determinateness by two combinations of spectral parameters from injectivity and interference observations. Even when accounting for the effects of finite length, storage capacity, skin, and non-parallelism of injection intervals and monitoring sections on the spectral parameters, it seems difficult to reconcile the bulk of our observations with the predictions of the GRF model. In cases, however, the gained fractal flow dimensions may bear geometrical significance, as for the likely anastomosing fault zone. The decrease in fractional flow regime with increasing mean interval pressure observed for two hydraulic paths between injection intervals and monitoring sections points towards a role of hydromechanical coupling in flow localization.
SummaryCertain geophysical inverse problems may be reduced to systems of sparse low-order polynomial equations. In simple systems of this type, the monomials are products of nodal values of a subsurface physical property, such as electrical conductivity, interior to a discretized modeling domain, whereas the coefficients are derived from the given boundary data. The latter are presumed to be over-specified by means of surface measurements made across a range of frequencies. Suppose an acceptable conductivity model is found, by any standard means, that fits the data. Then, provided other distinct, non-singular solutions exist, it is possible to find another one of them by charting a trajectory across model space using a path-tracking numerical method. A second path can then be tracked and, continuing in this manner, equivalent solutions to the inverse problem can be generated one by one. If the data are noisy, the equivalent models generate precisely the same numerical misfit. Regularization side constraints can be added which reduces the number of equivalent models and smooths the trajectories. This idea of tracking paths has been developed in the field of robotics where realizable configurations of kinematic mechanisms are sought; a similar approach is explored here, on small test systems, to explore its eventual applicability to practical geophysical inverse problems. In principle the method scales up to arbitrary size since the inverse of the forward matrix and the Jacobian can be expressed in polynomial form. The novel implications of a successful scaled-up implementation of this methodology is that multiple distinct inverse solutions can be obtained, each of which might entail a different geological interpretation. A surprising feature of the algorithm is that the forward problem becomes ill-posed while the inverse problem becomes well-posed.
A breakdown of normal polar winds in winter raises the chance of extreme cold—and a resulting spike in energy demand—in Europe, new research shows. Normally, strong and stable westerly winds circle high above the Arctic in the winter, in what is known as the "stratospheric polar vortex."
Just two weeks after it was launched, the Copernicus Sentinel-3C satellite returned its first images and measurements from two of its key instruments—a major milestone that confirms the satellite is working well and on track to monitor the health of Earth's oceans, land, cryosphere and atmosphere.
In high-latitude coastal areas, Arctic amplification means that temperatures are rising faster than the global average. This combination of rising temperatures and changing precipitation patterns driven by the North Atlantic Oscillation has major impacts on regions with a "cryosphere"—areas that host ice sheets and glaciers, such as southern Greenland.
Seismic data captured by ocean-bottom instruments have revealed "fossilized" traces of mantle flow beneath the Pacific Ocean's oldest preserved piece of crust. The patterns of ancient flow traced in the uppermost mantle are complicated, tracking changes in tectonic plate motion over time as well as mantle deformation around features such as a sunken remnant of the lithosphere (the crust and the uppermost brittle mantle) and upwelling hotspots.
Residents of Te Whanganui-a-Tara Wellington know it is always wise to prepare for any weather—to pack sunglasses, a raincoat and at least one extra layer. This "be prepared for anything" approach can equally be applied to Aotearoa's ocean weather during this year's "super" El Niño. Forecasts are fickle, unexpected outcomes are likely and preparation is needed.
Global climate and weather forecasting use predictive models that divide Earth's atmosphere into a grid, with each section typically covering tens to hundreds of kilometers. Now a new model, run by a team at the University of Tokyo, has narrowed this grid spacing to just 220 meters (720 feet). This higher resolution eliminated instances of "popcornlike" rain—bursts of heavy rain that appear in simulations but not in real life—providing more accurate predictions of extreme local weather from a global model.
Iron is one of the most abundant metals on Earth's surface, yet how such massive iron ore deposits formed throughout geological history has long remained a mystery. The prevailing theory has been that ancient photosynthetic bacteria oxidized iron, causing it to precipitate and accumulate in the oceans.
SummaryFull waveform inversion (FWI) is among the most accurate seismic imaging techniques, yet its Bayesian formulation remains computationally prohibitive. The gold-standard method, Markov chain Monte Carlo (McMC), requires hundreds of thousands of expensive forward simulations. Variational inference (VI) approaches offer improved efficiency but require gradient computations and may converge to incorrect posteriors. We introduce a probabilistic surrogate-assisted framework that replaces most expensive forward evaluations during sampling with predictions from a misfit-guided surrogate based on Deep Learning (DL) models, instantiated here using a convolutional neural network (CNN). The framework supports three regimes: (1) an offline regime in which the surrogate remains fixed after training, (2) an adaptive regime in which the surrogate is refined during sampling using a self-supervised active learning mechanism, and (3) a two-stage regime in which the surrogate screens proposals prior to exact evaluation. The two-stage regime preserves the exact posterior by construction, while the adaptive regime is asymptotically exact in the infinite-sample limit, provided that the surrogate approximation error vanishes. Because the target subsurface features in our study are spatially localised, we adopt a rock-physics-guided, shape-based parameterisation that represents each target by a small number of control nodes. This results in a low-dimensional inverse problem with a dimensionality of order 10 in the examples considered here, at which a misfit-guided DL surrogate becomes tractable. We demonstrate our framework on synthetic models of increasing dimensionality, including a CO2 plume monitoring scenario. The adaptive regime achieves a 7.1 × wall-clock speedup and reduces GPU-hour cost by approximately 84% relative to McMC with the exact forward solver, while recovering posteriors in close agreement with the reference. Comparison with Stein variational gradient descent (SVGD) shows the surrogate-assisted McMC sampling is approximately an order of magnitude more efficient in the examples considered here. This framework applies to any surrogate paired with any forward solver and objective function producing a scalar misfit.
SummarySeismic source classification often relies on a single best-fitting moment tensor and deterministic rules on source-type diagrams. This study formulates classification directly in terms of the posterior distribution of the moment tensor on the fundamental lune. Moment tensors from waveform inversion are mapped into a uniform (v, w) source-type domain and used to build empirical source-type probability density functions p(v, w). We evaluate five hypotheses (±ISOCAL, DC, ±CLVD) plus a residual baseline, where ISOCAL denotes a calibrated isotropic kernel relative to the canonical ISO endmember. Classification probabilities P(Hi) are obtained by integrating p(v, w) against the hypothesis kernels; a residual probability Presidual is included as a sixth baseline (uniform) model that absorbs probability when the posterior is diffuse or poorly matched by the prototypes. The framework is applied to published regional moment tensor ensembles for 32 events, including 23 nuclear tests at the Nevada National Security Site and the DPRK Punggye-Ri test site, plus nearby earthquakes and collapses. Well-recorded underground nuclear tests yield tightly localized +ISOCAL probabilities with small Presidual, while events recorded by sparse, complex source-receiver paths show broader probability vectors and larger Presidual. Tectonic earthquakes map primarily to double-couple and CLVD hypotheses, while collapses map primarily to negative isotropic (–ISOCAL), with secondary CLVD in some events. The method is applicable wherever source-type posteriors are available, given a declared hypothesis family, and produces explicit hypothesis probabilities with uncertainty and a residual diagnostic that flags diffuse or weakly matched posteriors, and can feed directly into monitoring and decision pipelines.
SummaryThe Marine Self-Potential (SP) method has proven to be a useful geophysical tool for localizing seafloor massive sulfide (SMS) deposits. Conventional geophysical inversion approaches face significant limitations in mid-ocean ridge environments with complex topography, including high computational requirements due to high degrees of freedom in grid discretization and the complex computation of sensitivities. Meanwhile, data-driven neural networks, while powerful, often suffer from poor generalizability and interpretability in complex deep-sea environments, owing to their reliance on large-scale training datasets. In this study, we propose a simple-formulated intelligent SP inversion framework, called SP-PINN, based on the Physics-Informed Neural Network (PINN). By embedding the governing equation of SP as a physical constraint into the loss function, SP-PINN effectively couples the intrinsic correlation between SP signals and subseafloor causative sources. To address the characteristics of topographic relief in mid-ocean ridges and discrete distribution of SMS ore bodies, 2D and 3D synthetic case studies were constructed to verify the inversion performance of SP-PINN. Two SMS deposit bodies on the Southwest Indian Ridge (SWIR) were successfully imaged and delineated through the application of SP-PINN to the field data. Our results demonstrate that accurate imaging of SMS can be achieved by introducing physical constraint terms into the neural network, properly optimizing hyperparameters, and incorporating a small amount of field data. This study provides a straightforward and physically interpretable SP imaging method for the assessment of SMS under data-scarce conditions.
SummaryShallow-water depth data serve as the essential foundation for marine spatial planning and coastal engineering projects. However, due to the attenuation of spectral signals, traditional retrieval of shallow-water seafloor topography struggles to maintain high accuracy in highly turbid waters and deeper areas. Furthermore, existing machine learning models incorporating spatial information are highly susceptible to spatial overfitting, which severely limits their generalization ability in complex waters. To address these issues, this study proposes a depth-stratified Extreme Gradient Boosting (XGBoost) approach combining Sentinel-2 remote sensing data and multi-source gravity data. This model innovatively incorporates gravity features, which are unaffected by water turbidity, effectively overcoming the signal attenuation limitations of spectral-only models in deep and turbid waters. In addition, the stratification strategy effectively mitigates spatial overfitting, which commonly affects machine learning models. Application results in the Pearl River Estuary, China, showed that compared to the unstratified spatially uncorrelated XGBoost global model, the root mean square error (RMSE) of the proposed model was significantly reduced by 1.362 m. Ablation experiments further confirmed that incorporating gravity features reduced the RMSE in deeper-water areas (depth > 30 m) by 1.033 m, and dropped the RMSE in highly turbid waters from 1.548 m to 0.955 m. Finally, validation in an adjacent area showed that the model maintained useful local spatial generalization under similar environmental conditions, with an RMSE of 0.591 m in the 30–35 m interval.
Tiny photosynthetic organisms in the Baltic Sea can contribute substantially to the transport of carbon from surface waters to deeper parts of the sea. A new study from the Department of Ecology, Environment and Plant Sciences at Stockholm University shows that their ability to form colonies and aggregates is an important factor determining how much of their biomass sinks.
On Earth, everything is moving. The planet spins, the tectonic plates shift and the spin axis wobbles, as do the magnetic poles, which makes determining how things move a rather tricky proposition. For Rice University's Richard Gordon, this problem came up when he was trying to understand the history of the Hawaiian hot spot, the deep-mantle source of the volcanism that created the Hawaiian–Emperor chain of islands and underwater volcanoes.
