Seismic velocities in rocks are highly sensitive to changes in permanent deformation and fluid content. The temporal variation of seismic velocity during the preparation phase of earthquakes has been well documented in laboratories but rarely observed in nature. It has been recently found that some anthropogenic, high-frequency (>1 Hz) seismic sources are powerful enough to generate body waves that travel down to a few kilometers and can be used to monitor fault zones at seismogenic depth. Anthropogenic seismic sources typically have fixed spatial distribution and provide new perspectives for velocity monitoring. In this work, we propose a systematic workflow to seek such powerful seismic sources in a rapid and straightforward manner. We tackle the problem from a statistical point of view, considering that persistent, powerful seismic sources yield highly coherent correlation functions (CFs) between pairs of seismic sensors. The algorithm is tested in California and Japan. Multiple sites close to fault zones show high-frequency CFs stable for an extended period of time. These findings have great potential for monitoring fault zones, including the San Jacinto Fault and the Ridgecrest area in Southern California, Napa in Northern California, and faults in central Japan. However, extra steps, such as beamforming or polarization analysis, are required to determine the dominant seismic sources and study the source characteristics, which are crucial to interpreting the velocity monitoring results. Train tremors identified by the present approach have been successfully used for seismic velocity monitoring of the San Jacinto Fault in previous studies.

A series of discrete elements simulations is presented for the study of fault gouges' frictional response. The gouge is considered to have previously undergone ultra-cataclastic flow and long-time consolidation loading. We explore the effect of different particle characteristics such as size, polydispersity, and also shearing velocities on gouge's response under the conditions met in the seismogenic zone. Monte-Carlo analyses suggest that the local stick-slip events disappear when averaging over a large number of numerical samples. Moreover, the apparent material frictional response remains almost unaffected by the spatial randomness of particles' position and by the particle's size distribution. On the contrary, the mean particle size controls the formation and thickness of the observed shear bands, which appear after the peak friction is met. Furthermore, the apparent friction evolution fits well to an exponential decay law with slip, which involves a particle size dependent critical slip distance. For the studied conditions and depth, the shearing velocity is found to play a secondary role on the apparent frictional response of the gouge, which highlights the importance of analyses involving multiphysics for studying the rheology of fault gouges. Besides improving the understanding of the underlying physics of the problem, the above findings are also useful for deriving pertinent constitutive models in the case of modeling with continuum theories.

Our understanding of geomagnetic field intensity prior to the era of direct instrumental measurements relies on paleointensity analysis of rocks and archaeological materials that serve as magnetic recorders. Only in rare cases are absolute paleointensity data sets continuous over millennial timescales, in sub-centennial resolution, and directly dated using radiocarbon. As a result, fundamental properties of the geomagnetic field, such as its maximum intensity and rate of change have remained a subject of lively discussion. Here, we place firm constraints on these two quantities using Bayesian modeling of well-dated archaeomagnetic intensity data from the Levant and Upper Mesopotamia. We report new data from 23 groups of pottery collected from 18 consecutive radiocarbon-dated archaeological strata from Tel Megiddo, Israel. In the Near East, the period of 1700-550 BCE is represented by 84 groups of archaeological artifacts, 55 of which were dated using radiocarbon or a direct link to clear historically dated events, providing unprecedented sub-century resolution. Moreover, stratigraphic relationships between samples collected from multi-layered sites enable further refinement of the data ages. The Bayesian curve shows four geomagnetic spikes between 1050 and 600 BCE, with virtual axial dipole moment (VADM) reaching values of 155-162 ZAm, much higher than any prediction from geomagnetic field models. Rates of change associated with the four spikes are ∼0.35-0.55 μT/year (∼0.7-1.1 ZAm/year), at least twice the maximum rate inferred from direct observations spanning the past 180 years. The increase from 1750 to 1030 BCE (73-161 ZAm) depicts the Holocene's largest change in field intensity.

FeO represents an important end-member for planetary interiors mineralogy. However, its properties in the liquid state under high pressure are poorly constrained. Here, in situ high-pressure and high-temperature X-ray diffraction experiments, ab initio simulations, and thermodynamic calculations are combined to study the local structure and density evolution of liquid FeO under extreme conditions. Our results highlight a strong shortening of the Fe-Fe distance, particularly pronounced between ambient pressure and ∼40 GPa, possibly related with the insulator to metal transition occurring in solid FeO over a similar pressure range. Liquid density is smoothly evolving between 60 and 150 GPa from values calculated for magnetic liquid to those calculated for non-magnetic liquid, compatibly with a continuous spin crossover in liquid FeO. The present findings support the potential decorrelation between insulator/metal transition and the high-spin to low-spin continuous transition, and relate the changes in the microscopic structure with macroscopic properties, such as the closure of the Fe-FeO miscibility gap. Finally, these results are used to construct a parameterized thermal equation of state for liquid FeO providing densities up to pressure and temperature conditions expected at the Earth's core-mantle boundary.

Enhanced earthquake catalogs provide detailed images of evolving seismic sequences. Currently, these data sets take some time to be released but will soon become available in real time. Here, we explore whether and how enhanced seismic catalogs feeding into established short-term earthquake forecasting protocols may result in higher predictive skill. We consider three enhanced catalogs for the 2016-2017 Central Italy sequence, featuring a bulk completeness lower by at least two magnitude units compared to the real-time catalog and an improved hypocentral resolution. We use them to inform a set of physical Coulomb Rate-and-State (CRS) and statistical Epidemic-Type Aftershock Sequence (ETAS) models to forecast the space-time occurrence of M3+ events during the first 6 months of the sequence. We track model performance using standard likelihood-based metrics and compare their skill against the best-performing CRS and ETAS models among those developed with the real-time catalog. We find that while the incorporation of the triggering contributions from new small magnitude detections of the enhanced catalogs is beneficial for both types of forecasts, these models do not significantly outperform their respective near real-time benchmarks. To explore the reasons behind this result, we perform targeted sensitivity tests that show how (a) the typical spatial discretizations of forecast experiments ( 2 km) hamper the ability of models to capture highly localized secondary triggering patterns and (b) differences in earthquake parameters (i.e., magnitude and hypocenters) reported in different catalogs can affect forecast evaluation. These findings will contribute toward improving forecast model design and evaluation strategies for next-generation seismic catalogs.

Accurate earthquake location and magnitude estimation play critical roles in seismology. Recent deep learning frameworks have produced encouraging results on various seismological tasks (e.g., earthquake detection, phase picking, seismic classification, and earthquake early warning). Many existing machine learning earthquake location methods utilize waveform information from a single station. However, multiple stations contain more complete information for earthquake source characterization. Inspired by recent successes in applying graph neural networks (GNNs) in graph-structured data, we develop a Spatiotemporal Graph Neural Network (STGNN) for estimating earthquake locations and magnitudes. Our graph neural network leverages geographical and waveform information from multiple stations to construct graphs automatically and dynamically by adaptive message passing based on graphs' edges. Using a recent graph neural network and a fully convolutional neural network as baselines, we apply STGNN to earthquakes recorded by the Southern California Seismic Network from 2000 to 2019 and earthquakes collected in Oklahoma from 2014 to 2015. STGNN yields more accurate earthquake locations than those obtained by the baseline models and performs comparably in terms of depth and magnitude prediction, though the ability to predict depth and magnitude remains weak for all tested models. Our work demonstrates the potential of using GNNs and multiple stations for better automatic estimation of earthquake epicenters.

High-resolution computer simulations of earthquake sequences in three or even two dimensions pose great demands on time and energy, making lower-cost simplifications a competitive alternative. We systematically study the advantages and limitations of simplifications that eliminate spatial dimensions in quasi-dynamic earthquake sequence models, from 3D models with a 2D fault plane down to 0D or 1D models with a 0D fault point. We demonstrate that, when 2D or 3D models produce quasi-periodic characteristic earthquakes, their behavior is qualitatively similar to lower-dimension models. Certain coseismic characteristics like stress drop and fracture energy are largely controlled by frictional parameters and are thus largely comparable. However, other observations are quantitatively clearly affected by dimension reduction. We find corresponding increases in recurrence interval, coseismic slip, peak slip velocity, and rupture speed. These changes are to a large extent explained by the elimination of velocity-strengthening patches that transmit tectonic loading onto the velocity-weakening fault patch, thereby reducing the interseismic stress rate and enhancing the slip deficit. This explanation is supported by a concise theoretical framework, which explains some of these findings quantitatively and effectively estimates recurrence interval and slip. Through accounting for an equivalent stressing rate at the nucleation size * into 2D and 3D models, 0D or 1D models can also effectively simulate these earthquake cycle parameters. Given the computational efficiency of lower-dimensional models that run more than a million times faster, this paper aims to provide qualitative and quantitative guidance on economical model design and interpretation of modeling studies.

Many described subduction complexes (or mélanges) exhumed from seismogenic depths comprise thick, turbidite-dominated sequences with deformed zones containing clasts or boudins of more competent sandstone and/or basalt. In contrast, many active subduction zones have a relatively small thickness of sedimentary inputs (<2 km), turbidite sequences are commonly accreted rather than subducted, and the role of pelagic sediments and basalt (lavas and hyaloclastites) in the deforming zone near the plate interface at <20 km depth is poorly understood. Field investigation of Neoproterozoic oceanic sequences accreted in the Gwna Complex, Anglesey, UK, reveals repeated lenticular slices of variably sampled ocean plate stratigraphy (OPS) bounded by thin mélange-bearing shear zones. Mélange matrix material is derived from adjacent OPS lithologies and is either dominantly illitic, likely derived from altered siliciclastic sediment, or chloritic, likely derived from altered volcanics. In the illitic mélange, mutually cross-cutting phyllosilicate foliation and variably deformed chlorite-quartz-calcite veins suggest ductile creep was cyclically punctuated by transient, localized fluid pulses. Chlorite thermometry indicates the veins formed at 260 ± 10°C. In the chloritic mélange, recrystallized through-going calcite veins are deformed to shear strains of 4-5 within a foliated chlorite matrix, suggesting calcite veins in subducting volcanics may localize deformation in the seismogenic zone. Shear stress-strain rate curves constructed using existing empirical relationships in a simplified shear zone geometry predict that slip velocities varied depending on pore fluid pressure; models predict slow slip velocities preferentially by frictional sliding in chlorite, at pore fluid pressures greater than hydrostatic but less than lithostatic.

Fault surfaces are rough at all scales, and this significantly affects fault-slip behavior. However, roughness is only occasionally considered experimentally and then often in experiments imposing a low-slip velocity, corresponding to the initiation stage of the earthquake cycle. Here, the effect of roughness on earthquake nucleation up to runaway slip is investigated through a series of dry load-stepping biaxial experiments performed on bare rock surfaces with a variety of roughnesses. These laboratory faults reached slip velocities of at least 100 mm/s. Acoustic emissions were located during deformation on bare rock surfaces in a biaxial apparatus during load-stepping experiments for the first time. Smooth surfaces showed more frequent slip instabilities accompanied by slip bursts and larger stress drops than rough faults. Smooth surfaces reached higher slip velocities and were less inclined to display velocity-strengthening behavior. The recorded and localized acoustic emissions were characterized by a greater proportion of large-magnitude events, and therefore likely a higher Gutenberg-Richter -value, for smoother samples, while the cumulative seismic moment was similar for all roughnesses. These experiments shed light on how local microscopic heterogeneity associated with surface topography can influence the macroscopic stability of frictional interfaces and the associated microseismicity. They further provide a laboratory demonstration of roughness' ability to induce stress barriers, which can halt rupture, a phenomenon previously shown numerically.

Understanding the stress evolution of extinct volcanoes can improve efforts to forecast flank eruptions on active systems. Field, petrographic, and seismic data are combined with numerical modeling to investigate the paleo-stress field of New Zealand's Akaroa Volcano, or Akaroa Volcanic Complex. Field mapping identifies 86 radially oriented dikes and seven lava domes found only within a narrow elevation range along Akaroa's erosional crater rim. These observations suggest that crater rim dike emplacement resulted from lateral deflection of vertically ascending intrusions from a centralized magma source, which in turn may have facilitated formation of the lava domes, as well as two scoria cones. We postulate that dike deflection occurred along a stress barrier, as neither a compositional change nor structural boundary are present. We use a finite element model (FEM) simulating Akaroa to test how different factors may have influenced the system's stress state and dike geometry. Elastic, non-flexural ("roller") model configurations containing a large, oblate, and shallow magma chamber produce stress barriers most conducive to radial dike emplacement along Akaroa's crater rim. These configurations also simulate rapid edifice construction above a preexisting lithospheric "bulge." Conversely, simulating flexural stresses exerted on the lithosphere by Akaroa's large mass hinder rather than promote radial dike emplacement. Temperature-dependent viscoelastic relaxation promotes gradual increases in stress barrier elevation, though this effect is strongly dependent on magma chamber parameters. These results suggest that Akaroa was constructed rapidly (within ∼100 kyr) prior to crater rim dike emplacement, which occurred throughout the volcano's remaining active lifespan.

We propose a theory for preventing instabilities and inducing controlled, slow-slip in frictionally unstable systems, such as the Generalized-Burridge-Knopoff (GBK) model and seismic fault models. We exploit the dependence of friction on pressure and use it as a backdoor for altering the dynamics of the underlying dynamical system. We use the mathematical Theory of Control and, for the first time, we manage to (a) stabilize and restrict chaos in this kind of systems, (b) guarantee slow frictional dissipation and (c) tune the system toward desirable global asymptotic equilibria of lower energy. Our control approach is robust and does not require exact knowledge of the frictional or elastic behavior of the system. Numerical examples of control are given for a Burridge-Knopoff system and a strike-slip fault model obeying rate-and-state friction. GBK models are known to present Self-Organized Critical (SOC) behavior. Therefore, the presented methodology shows an additional example of SOC Control. Even though further developments are necessary before any practical application, we expect our methodology to inspire earthquake mitigation strategies regarding anthropogenic and/or natural seismicity.

Aseismic afterslip is postseismic fault sliding that may significantly redistribute crustal stresses and drive aftershock sequences. Afterslip is typically modeled through geodetic observations of surface deformation on a case-by-case basis, thus questions of how and why the afterslip moment varies between earthquakes remain largely unaddressed. We compile 148 afterslip studies following 53 6.0-9.1 earthquakes, and formally analyze a subset of 88 well-constrained kinematic models. Afterslip and coseismic moments scale near-linearly, with a median Spearman's rank correlation coefficient (CC) of 0.91 after bootstrapping (95% range: 0.89-0.93). We infer that afterslip area and average slip scale with coseismic moment as and , respectively. The ratio of afterslip to coseismic moment ( ) varies from <1% to >300% (interquartile range: 9%-32%). weakly correlates with (CC: -0.21, attributed to a publication bias), rupture aspect ratio (CC: -0.31), and fault slip rate (CC: 0.26, treated as a proxy for fault maturity), indicating that these factors affect afterslip. does not correlate with mainshock dip, rake, or depth. Given the power-law decay of afterslip, we expected studies that started earlier and spanned longer timescales to capture more afterslip, but does not correlate with observation start time or duration. Because estimates for a single earthquake can vary by an order of magnitude, we propose that modeling uncertainty currently presents a challenge for systematic afterslip analysis. Standardizing modeling practices may improve model comparability, and eventually allow for predictive afterslip models that account for mainshock and fault zone factors to be incorporated into aftershock hazard models.

Global Navigation Satellite System (GNSS) vertical displacements measuring the elastic response of Earth's crust to changes in hydrologic mass have been used to produce terrestrial water storage change (∆TWS) estimates for studying both annual ∆TWS as well as multi-year trends. However, these estimates require a high observation station density and minimal contamination by nonhydrologic deformation sources. The Gravity Recovery and Climate Experiment (GRACE) is another satellite-based measurement system that can be used to measure regional TWS fluctuations. The satellites provide highly accurate ∆TWS estimates with global coverage but have a low spatial resolution of ∼400 km. Here, we put forward the mathematical framework for a joint inversion of GNSS vertical displacement time series with GRACE ∆TWS to produce more accurate spatiotemporal maps of ∆TWS, accounting for the observation errors, data gaps, and nonhydrologic signals. We aim to utilize the regional sensitivity to ∆TWS provided by GRACE mascon solutions with higher spatial resolution provided by GNSS observations. Our approach utilizes a continuous wavelet transform to decompose signals into their building blocks and separately invert for long-term and short-term mass variations. This allows us to preserve trends, annual, interannual, and multi-year changes in TWS that were previously challenging to capture by satellite-based measurement systems or hydrological models, alone. We focus our study in California, USA, which has a dense GNSS network and where recurrent, intense droughts put pressure on freshwater supplies. We highlight the advantages of our joint inversion results for a tectonically active study region by comparing them against inversion results that use only GNSS vertical deformation as well as with maps of ∆TWS from hydrological models and other GRACE solutions. We find that our joint inversion framework results in a solution that is regionally consistent with the GRACE ∆TWS solutions at different temporal scales but has an increased spatial resolution that allows us to differentiate between regions of high and low mass change better than using GRACE alone.

Ab initio molecular dynamics calculations on a carbonate-silicate-metal melt were performed to study speciation and coordination changes as a function of pressure and temperature. We examine in detail the bond abundances of specific element pairs and the distribution of coordination environments over conditions spanning Earth's present-day mantle. Average coordination numbers increase continuously from 4 to 8 for Fe and Mg, from 4 to 6 for Si, and from 2 to 4 for C from 1 to 148 GPa (4,000 K). Speciation across all pressure and temperature conditions is complex due to the unusual bonding of carbon. With the increasing pressure, C-C and C-Fe bonding increase significantly, resulting in the formation of carbon polymers, C-Fe clusters, and the loss of carbonate groups. The increased bonding of carbon with elements other than oxygen indicates that carbon begins to replace oxygen as an anion in the melt network. We evaluate our results in the context of diamond formation and of metal-silicate partitioning behavior of carbon. Our work has implications for properties of carbon and metal-bearing silicate melts, such as viscosity, electrical conductivity, and reactivity with surrounding phases.

The Gravity Recovery and Climate Experiment Follow-On (GRACE-FO), launched May 22, 2018 and collecting science data since June 2018, is extending the 15-year data record of Earth mass change established by its predecessor GRACE mission (2002-2017). The GRACE-FO satellites carry onboard a novel technology demonstration instrument for intersatellite ranging, the Laser Ranging Interferometer (LRI), in addition to the microwave interferometer (MWI) carried on GRACE. The LRI has out-performed its in-orbit performance requirements both in terms of accuracy as well as the duration of tracking. Here, we compare and validate LRI-based gravity solutions for January 2019 to September 2020 against the MWI solutions. The comparison between the two sets of gravity solutions shows great similarities in general and nearly perfect consistency at a large hydrologic basin spatial scale (100,000 km and above), commonly viewed as the spatial resolution established by GRACE. The comparison in the spectral domain shows differences at the higher degrees of the spectrum, with lower error in the zonal and near zonal terms for the LRI solutions. We conclude that the LRI observations can be used to recover time-varying gravity signals to at least the level of accuracy established by the MWI-based solutions. This is a promising finding, especially when considering the benefits of using the LRI over the MWI, such as the great stability of the instrument and the low occurrence of instrument reboot events.

We present a novel methodology for exploring 4D seismic data in the context of monitoring subsurface resources. Data-space exploration is a key activity in scientific research, but it has long been overlooked in favor of model-space investigations. Our methodology performs a data-space exploration that aims to define structures in the covariance matrix of the observational errors. It is based on Bayesian inferences, where the posterior probability distribution is reconstructed through trans-dimensional (trans-D) Markov chain Monte Carlo sampling. The trans-D approach applied to data-structures (termed "partitions") of the covariance matrix allows the number of partitions to freely vary in a fixed range during the McMC sampling. Due to the trans-D approach, our methodology retrieves data-structures that are fully data-driven and not imposed by the user. We applied our methodology to 4D seismic data, generally used to extract information about the variations in the subsurface. In our study, we make use of real data that we collected in the laboratory, which allows us to simulate different acquisition geometries and different reservoir conditions. Our approach is able to define and discriminate different sources of noise in 4D seismic data, enabling a data-driven evaluation of the quality (so-called "repeatability") of the 4D seismic survey. We find that: (a) trans-D sampling can be effective in defining data-driven data-space structures; (b) our methodology can be used to discriminate between different families of data-structures created from different noise sources. Coupling our methodology to standard model-space investigations, we can validate physical hypothesis on the monitored geo-resources.

Lithospheric net rotation (LNR) is the movement of the lithosphere as a solid body with respect to the mantle. Separating the signal of LNR from plate tectonic motion is therefore an important factor in producing absolute plate motion models. Net rotation is difficult to constrain because of uncertainties in geological data and outstanding questions about the stability of the mantle plumes used as a reference frame. We use mantle convection simulations to investigate the controlling factors for the magnitude of LNR and to find the statistical predictability of LNR in a fully self-consistent convective system. We find that high lateral viscosity variations are required to produce Earth-like values of LNR. When the temperature dependence of viscosity is lower, and therefore slabs are softer, other factors such as the presence of continents and a viscosity gradient at the transition zone are also important for determining the magnitude of net rotation. We find that, as an emergent property of the chaotic mantle convection system, the evolution of LNR is too complicated to predict in our models. However, we find that the range of LNR within the simulations follows a Gaussian distribution, with a correlation time of 5 Myr. The LNR from the models needs to be sampled for around 50 Myr to produce a fully Gaussian distribution. This implies, that within the time frames considered for absolute plate motion reconstructions, LNR can be treated as a Gaussian variable. This provides a new geodynamic constraint for absolute plate motion reconstructions.

Our understanding of the past behavior of the geomagnetic field arises from magnetic signals stored in geological materials, e.g., (volcanic) rocks. Bulk rock samples, however, often contain magnetic grains that differ in chemistry, size, and shape; some of them record the Earth's magnetic field well, others are unreliable. The presence of a small amount of adverse behaved magnetic grains in a sample may already obscure important information on the past state of the geomagnetic field. Recently it was shown that it is possible to determine magnetizations of individual grains in a sample by combining X-ray computed tomography and magnetic surface scanning measurements. Here we establish this new Micromagnetic Tomography (MMT) technique and make it suitable for use with different magnetic scanning techniques, and for both synthetic and natural samples. We acquired reliable magnetic directions by selecting subsets of grains in a synthetic sample, and we obtained rock-magnetic information of individual grains in a volcanic sample. This illustrates that MMT opens up entirely new venues of paleomagnetic and rock-magnetic research. MMT's unique ability to determine the magnetization of individual grains in a nondestructive way allows for a systematic analysis of how geological materials record and retain information on the past state of the Earth's magnetic field. Moreover, by interpreting only the contributions of known magnetically well-behaved grains in a sample, MMT has the potential to unlock paleomagnetic information from even the most complex, crucial, or valuable recorders that current methods are unable to recover.

Basaltic and gabbroic rocks that define the seafloor have been suggested to act as sources of rheological heterogeneity during subduction, with the capacity to enhance or dampen seismicity. Despite this, relatively little is known from the rock record regarding the progression and conditions of mafic oceanic crust deformation during subduction, particularly in the shallow megathrust region of the seismogenic zone. We describe subduction-related deformation structures and characterize deformation conditions from an exhumed, basalt-hosted megathrust in the Chugach accretionary complex of south-central Alaska. Rocks in the Chugach preserve a record of seafloor mineralogical changes from pre-subduction, hydrothermal circulation that produced sheet silicates with a lower frictional strength than intact basalt. Pre-subduction alteration also served to introduce hydrous phases that can expel water during deformation and raise the pore fluid pressure. Once strain localized within basalts onto a megathrust fault plane at lithostatic pore fluid pressures, the basalt weakened further through a combination of cataclasis, dilatational shear fracturing, and slip on chlorite-rich shear bands. This process occurred in a narrower fault zone, and at higher maximum differential stress and greater pore fluid pressure fluctuations than recorded in some sediment-hosted megathrusts at similar pressure and temperature conditions. Our data indicate that when the lower plate contains basalt bathymetric features, basalt dismembers during subduction into a chlorite-rich fault gouge that surrounds lenses or slices of intact, less-altered basalt.

Machine learning (ML) techniques have become increasingly important in seismology and earthquake science. Lab-based studies have used acoustic emission data to predict time-to-failure and stress state, and in a few cases, the same approach has been used for field data. However, the underlying physical mechanisms that allow lab earthquake prediction and seismic forecasting remain poorly resolved. Here, we address this knowledge gap by coupling active-source seismic data, which probe asperity-scale processes, with ML methods. We show that elastic waves passing through the lab fault zone contain information that can predict the full spectrum of labquakes from slow slip instabilities to highly aperiodic events. The ML methods utilize systematic changes in P-wave amplitude and velocity to accurately predict the timing and shear stress during labquakes. The ML predictions improve in accuracy closer to fault failure, demonstrating that the predictive power of the ultrasonic signals improves as the fault approaches failure. Our results demonstrate that the relationship between the ultrasonic parameters and fault slip rate, and in turn, the systematically evolving real area of contact and asperity stiffness allow the gradient boosting algorithm to "learn" about the state of the fault and its proximity to failure. Broadly, our results demonstrate the utility of physics-informed ML in forecasting the imminence of fault slip at the laboratory scale, which may have important implications for earthquake mechanics in nature.

We present seismic tomographic results from a unique seismic refraction and wide-angle survey along a 600 km long flow-line corridor of oceanic lithosphere ranging in age from 0 to 27 Ma in the equatorial Atlantic Ocean at 2° 43'S. The velocities in the crust near the ridge axis rapidly increase in the first 6 Myr and then change gradually with age. The upper crust (Layer 2) thickness varies between 2 and 2.4 km with an average thickness of 2.2 km and the crustal thickness varies from 5.6 to 6 km along the profile with an average crustal thickness of 5.8 km. At some locations, we observe negative velocity anomalies (∼-0.3 km/s) in the lower crust which could be either due to chemical heterogeneity in gabbroic rocks and/or the effects of fault related deformation zones leading to an increase in porosities up to 1.6% depending on the pore/crack geometry. The existence of a low velocity anomaly beneath the ridge axis suggests the presence of partial melt (∼1.3%) in the lower crust. Upper mantle velocities also remain low (∼7.8 km/s) from ridge axis up to 5 Ma, indicating a high temperature regime associated with mantle melting zone underneath. These results suggest that the evolution of the crust and uppermost mantle at this location occur in the first 10 Ma of its formation and then remains unchanged. Most of the structures in the older crust and upper mantle are fossilized structures and could provide information about past processes at ocean spreading centers.