Geophysical Journal International

SummaryThe Orientale Basin is the youngest and best-preserved multi-ring impact basin on the Moon. It is approximately 930 km in diameter and comprises three concentric rings—the Cordillera, Inner Rook, and Outer Rook rings. We used Gravity Recovery and Interior Laboratory (GRAIL) mission gravity data (GRGM1200B, truncated to degree 660) to invert the three-dimensional (3D) density structure associated with the basin’s mascon and ring-related crustal anomalies. To separate longer and shorter-wavelength signals, we performed inversions using (1) spherical harmonic degrees 2–660 to characterise the basin’s deep structure and (2) degrees 60–660 to highlight crustal-scale heterogeneity. The inversion with degree and order of 2–660 resolves a basin-wide central positive density anomaly beneath the inner depression, extending from a depth of ∼16–80 km, corresponding to uplifted mantle. This anomaly persists below the crust–mantle boundary; however, its deep continuation should be interpreted cautiously because it may partially reflect vertical smearing in the gravity inversion. Nonetheless, the spatial association of this anomaly with a mascon rooted in the upper mantle is compatible with impact-driven uplift of dense material from depth, followed by post-impact thermal evolution and relaxation processes. The inversion with degree and order of 60–660 indicates alternating positive and negative density anomalies associated with the ring system. Prominent ring-parallel positive anomalies occur along the Outer Rook Ring and Cordillera Ring, extending to ∼30–35 km depth and exhibiting a density contrast of ∼60–200 kg/m3. The geometry and lateral continuity of these anomalies across multiple rings support an interpretation of ring-controlled crustal heterogeneity, consistent with either intrusion and structurally focused modification along ring-related discontinuities or impact-generated fracturing that provided pathways for magma ascent. The results of this study provide quantitative constraints on the depth, magnitude, and spatial distribution of density anomalies associated with the Orientale mascon and ring system, thereby improving the subsurface framework for regional geological interpretation and supporting future lunar landing-site geophysical investigations and in situ sampling. Based on the inferred crustal density architecture, we propose that future geophysical sampling efforts should prioritise the Outer Rook Ring and Cordillera Ring, where the observed ring-parallel anomalies may provide insights into crustal modification processes and the composition of the lunar interior.

SummaryWe studied ionospheric changes associated with the 2025 March 28 Myanmar earthquake (Mw7.7) using global navigation satellite system receivers to measure ionospheric electrons, as a part of the project to predict earthquake precursors. The total electron contents above the fault changed their trends ~36 minutes before the earthquake, with the positive anomaly reaching ~1 per cent of the background. These quantities fit well with the past ~20 cases despite relatively large day-to-day variability due to high geomagnetic activities. The positive anomaly was sandwiched by two negative anomalies to the north and the south, suggesting within-ionosphere electron transportation along geomagnetic fields possibly driven by surface positive electric charges released from the fault.

SummaryWe develop and demonstrate a new paradigm for modeling prompt forensics data from potential nuclear explosions of concern. Related scenarios include nuclear terrorism, which may involve low-yield detonations. Traditional modeling appropriate for higher-yield historical nuclear testing is generalized to capture uncertainties in yields (such as when using conversion to obtain “equivalent” nuclear yields for conventional explosives in low-yield experiments) and to capture variation among source-to-sensor path effects for which no calibration data are available. Special attention is paid to quantifying these sources of uncertainty and to their formal inclusion in comprehensive yield uncertainty — uncertainty that would otherwise be underestimated and potentially lead to mistaken conclusions. For the example scenario considered, two useful stand-alone monitoring phenomenologies are based on geophysical data (from surface effects characteristics and local seismic metrics). By fusing signatures from multiple phenomenologies, the Multi-Phenomenology Explosion Monitoring (MultiPEM) framework provides improved yield characterization relative to reliance on single-phenomenology analysis alone, especially when individual sensors have complementary sensitivities to emplacement/environmental conditions.

SummaryBroadband seismic stations are primarily designed to record ground displacement from earthquakes, but they are sensitive to a wide range of processes, including human activity, oceanic waves, and atmospheric pressure variations. These signals are often considered noise, yet their study at frequencies from a few millihertz to one hertz has been fundamental for understanding geosphere coupling, developing methods to image Earth’s interior, and monitoring climate. At lower frequencies, below 10 cycles per day, the origin of continuous seismic noise remains poorly understood and may result from multiple coexisting mechanisms. To illuminate this part of the spectrum and its governing physics, we apply a dedicated processing method to 20 years of global seismic data. Our approach enables precise quantification of frequencies, angular degrees, and velocities of low-frequency modes, which we unambiguously identify as Lamb waves in the atmosphere.

SummaryThe February 6, 2023, Mw 7.8 Kahramanmaraş earthquake reactivated the East Anatolian Fault Zone (EAFZ) and involved surface rupture from the Amik Basin in south to the northeast of Çelikhan in north. Although the EAFZ continues southwest towards the Gulf of İskenderun across the Amanos Mountain Range (AMR) via the Türkoğlu–Osmaniye segment, the 2023 rupture instead followed the western margin of the Karasu Valley. Our analysis suggests that this deviation was governed by structural architecture of these two routes. While both routes involve similar rock units, the AMR is characterized by a massive, intact, and thicker crust that allowed the velocity-strengthening frictional properties of the basement to act as an effective barrier, arresting the rupture east of Türkoğlu. In contrast, the transtensional architecture of the Karasu Valley—evidenced by deep-seated extensional fissures and basaltic volcanism—represents a structurally dilated and thinner crust. This extensional setting neutralized the potential frictional resistance of the basement, providing a rheologically compliant path of least resistance for the 2023 rupture. The Türkoğlu–Osmaniye segment has not experienced a major event for about last 1 500 years and considering a slip rate of ~4.8 mm/yr, about 7 meters of slip deficit has accumulated. Furthermore, the 2023 earthquake loaded stress on the Türkoğlu-Osmaniye Fault. For these reasons, the potential for an M > 7 earthquake on this part of the EAFZ is high, necessitating urgent preparedness for a large earthquake in the region.

SummaryThe Born approximation offers a computationally efficient alternative to full electromagnetic (EM) forward modeling, but suffers from limited accuracy due to its reliance on a fixed background conductivity. In this work, we develop an adaptive Born approximation that treats the background medium as a tunable parameter to enhance accuracy in a goal-oriented manner. The background conductivity is selected locally for each measurement configuration using spatial sensitivity functions, enabling accurate modeling in both isotropic and anisotropic media. In this study, we primarily focus on horizontally layered earth models penetrated by a vertical well to investigate the fundamental behavior of the approximation in a simplified setting. We formulate our approach to be applicable to general anisotropic media by using the Green’s function defined for a homogeneous medium. Furthermore, the approach extends to cases where the background conductivity is isotropic while the actual medium is anisotropic. For a layered medium, the orientation of induced current densities relative to the layering provides physical intuition for background selection, drawing analogies to Voigt- and Reuss-type bounds. While these analogies offer useful guidance, our numerical results do not always conform to the expectations derived from them. Among the averaging schemes evaluated, arithmetic averaging generally yields the most accurate results. Numerical experiments indicate that the adaptive approach significantly outperforms fixed-background models across a range of frequencies, spacings, and conductivity contrasts. Furthermore, an example with a 3D structure illustrates the method’s broader applicability beyond the horizontally-layered earth setting. This framework provides a principled and efficient path toward fast, accurate EM borehole modeling for real-time well geosteering and subsurface electrical imaging.

SummaryThe Southern Apennines—Northern Calabrian boundary is a region marked by lithological heterogeneity, complex geodynamics and tectonics, and prone to significant seismic hazard. This sector is part of a complex geodynamic system, where Africa-Eurasia convergence, Ionian subduction, and slab retreat coexist. Its structure and seismic activity derive from extensive lithospheric heterogeneity and fluid-related processes, both of which are poorly constrained. Here, we present a novel application of seismic attenuation and scattering tomography of the area at a regional scale. We estimated seismic wave attenuation and scattering for the Southern Apennines—Northern Calabria region using a dataset of 1581 waveforms related to 95 M ≥ 3.0 earthquakes that occurred between 2004 and 2024 and were recorded at 32 stations. We constrained the heterogeneous properties and fluid saturation of the Southern Apennines—Northern Calabrian region by mapping P-wave Peak Delays and inverting coda-normalized energies for total attenuation (1/Q). Results consistently reveal different seismic energy dissipation mechanisms between the two domains, reflecting their different characteristics in terms of Peak Delay and attenuation patterns. The Southern Apennines exhibit high Peak Delay values at all depths and almost no remarkable total attenuation anomalies, consistent with weakly consolidated, fractured sedimentary sequences and limited fluid content. Nevertheless, at a depth of 5.4 km, a relatively high attenuation pattern is detectable, likely linked to the presence of less cohesive and potentially fluid-saturated units. Conversely, Northern Calabria shows low Peak Delay and high attenuation in the investigated depth range, reflecting wave propagation through coherent crystalline rocks with significant fluid circulation, likely favored by overpressurized materials or active migration pathways. The spatial correlation between high attenuation, low seismic velocities, and thermal anomalies shows that fluids modulate seismic wave behavior, providing new constraints on the crustal structure and seismotectonic segmentation of the region. The joint interpretation of our results with other geophysical models and responses highlights the complex interplay between lithology, tectonics, and fluid dynamics across this critical segment of the central Mediterranean.

SummaryThe earthquake fault as observed by seismic motion primarily manifests as a surface of displacement discontinuity within a linear elastic continuum. The displacement discontinuity and the surface normal vector (n-vector) of this idealized earthquake source are measured by the tensor of potency, which is seismic moment normalized by stiffness. We exploit this theoretical relation to formulate an inverse problem of reconstructing a smooth, three-dimensional fault surface from an areal density field of the potency tensor. In this problem, the surface is represented by an elevation field that parametrizes the vertical variation of the surface relative to a reference, and the nodal planes of a given potency-density-tensor field describe the n-vector field. The remaining subject is the n-vector-to-elevation transform, the operation inverse to defining the n-vector field on a given surface. Whereas this transform is a well-posed one-to-one mapping in two dimensions where the n-vector has one degree of freedom, the transform becomes overdetermined in three dimensions because the n-vector has two degrees of freedom while the scalar elevation has only one, generally admitting no solution. This overdetermination originates from a reduction in degrees of freedom from six to five upon modeling the source as a displacement discontinuity rather than general potency density, namely inelastic strain. The sixth degree of freedom unmodeled by displacement discontinuities and n-vectors manifests as a local violation of the determinant-free constraint in point potency sources; however, in areal sources of potency density, it raises a conflict with the global consistency of the n-vector field. Recognizing that this conflict derives from the capacity of the potency-density-tensor field to describe inelastic strain source incompatible with displacement discontinuity on a surface, we explicitly introduce an a priori constraint to define the fault surface as the smooth surface that best approximates the surface distribution of inelastic strain by displacement discontinuity. We derive an analytical solution for the surface reconstruction thus formulated and demonstrate its ability to reproduce smooth three-dimensional surfaces from synthetic noisy n-vector fields. Lastly, we integrate the derived formula into the potency density tensor inversion and validate it in an application to the 2013 Balochistan earthquake. The estimated fault geometry agrees better with the observed fault trace than that of the previously proposed quasi-two-dimensional surface reconstruction, highlighting the importance of accounting for three-dimensional fault geometry.

SummarySimulating the behavior of geological materials represents a fundamental objective in geophysical research. To achieve this goal, various models have been developed for different scenarios. While continuum models based on continuum mechanics are most commonly employed, non-continuum approaches such as the discrete element model and the lattice model have also been developed to address the pervasive discontinuities inherent in geological materials. However, existing non-continuum models are predominantly limited to isotropic conditions, significantly constraining their applicability. The dynamic lattice method proposed in this study aims to overcome this limitation. By independently assigning elastic properties to lattice bonds based on their spatial orientation, we have successfully introduced anisotropy into a three-dimensional lattice model. The linear relationships between the lattice model parameters and their continuum counterparts are established in terms of elastic properties. This advanced three-dimensional lattice model has been effectively applied in elastic seismic wave simulations in arbitrary anisotropic media.

SummaryThe singularity points are very common in the low-symmetry anisotropic media. The presence of these points results in complications of the wave fronts (or group velocity images) for waves forming the singularity points. We show the effect of singularity points in multilayered anisotropic media with triclinic symmetry.

SummaryQuantification and control of the in situ mechanics of calcite precipitation in the subsurface are notorious problems often encountered in hydrogeological and engineering applications. Difficulties arise here due to the general inaccessibility and texture of pore spaces, as well as the precipitation reaction’s dependence on fluid chemistry and the composition of the pore surfaces. To mitigate the uncertainties introduced by these inaccessible variables, we propose the use of spectral induced polarization (SIP) as a non-invasive tool to gain insight into the textural and electrochemical parameters controlling the precipitation rate within confined pore spaces and incorporate the gained information into a reactive transport model for quantification. We present SIP monitoring data from three laboratory experiments on diffusive mixing, inducing CaCO3 precipitation in sandstone. During the experimental runs, we identify a clear pattern showing the onset of the chemical reaction in the low-frequency (< 0.1 Hz) response of the imaginary conductivity and the formation of an associated high-frequency peak (> 10 Hz) in the later stages of the experiment. The changes to pore space geometry and precipitation yield were estimated with multiple independent methods. Using information gained from the monitoring data, we predict the dynamics of the precipitation reaction by including textural information, the inner surface area, and the grain size of the precipitate, as well as constraints on the effective diffusivity. These parameters were determined for each sample, based on empirical relations to the polarization response, and incorporated into an accompanying reactive transport model (phreeqc). The experimental results highlight the benefits SIP monitoring can provide to reactive transport models, even if precipitation yield is close to the detection limit of commonly applied methods, such as X-ray powder diffraction or fluorescence.

SummaryPrevious studies of the 2021 Mw 7.4 Maduo earthquake have primarily focused on the early postseismic phase, while the dominant mechanisms driving postseismic processes and the seismic moment released by afterslip remain debated. Longer-term observational constraints are needed to address these issues. In this study, we integrate ~3.5 years of postseismic InSAR and GPS time series data to effectively separate the contributions of afterslip and viscoelastic relaxation. The results show that afterslip released a cumulative seismic moment of approximately 3.91 × 1019 N·m, accounting for ~23.3 per cent of the coseismic moment—equivalent to a new Mw 7.0 earthquake. The optimal steady-state and transient viscosities of the lower crust are estimated to be 1.35 × 1019 Pa·s and 1.5 × 1018 Pa·s, respectively. Afterslip remains the dominant mechanism driving near-field deformation throughout the observation period, while viscoelastic relaxation governed far-field deformation beginning about 4 months after the mainshock. The stress-driven afterslip is comparable with the inverted kinematic afterslip, and poroelastic rebound is negligible. These findings provide valuable insights into stress perturbations on surrounding faults induced by the coseismic rupture, afterslip, and viscoelastic relaxation, and offer new constraints on the recurrence interval of Mw 7.4 earthquake on the Jiangcuo Fault.

SummaryMachine learning models offer powerful predictive capabilities for geoscientific applications but remain limited by their ”black-box” nature and lack of rigorous uncertainty quantification. We developed a comprehensive, generalizable uncertainty quantification framework that decomposes predictive uncertainty into aleatoric and epistemic components using Quantile Regression Forests. Additionally, we applied unsupervised k-means clustering to isolate homogeneous data regimes, thereby reducing aleatoric uncertainty across spatially heterogeneous geoscientific datasets. To facilitate interpretation and quality assessment, we introduced five spatial diagnostic tools: bandwidth, variance, robustness, confidence, and explainability maps that characterize prediction reliability and identify dominant uncertainty sources. To demonstrate the framework’s applicability, we tested it on three synthetic datasets varying in size and a real-world geothermal heat flow application with 14 geophysical observables across continental Africa. Results show that clustering substantially reduces aleatoric uncertainty while maintaining stable epistemic uncertainty. Clustering also improves predictive accuracy and sharpens prediction intervals, with gains most pronounced in homogeneous regions. Applied to the African geothermal heat flow, the framework reveals region-specific geological controls (lithospheric architecture dominates stable cratons, while tectonic proximity governs active rift zones) and guides targeted data collection by distinguishing high-epistemic regions requiring additional sampling from high-aleatoric zones needing improved observables. While theoretically applicable to other geographic regions and geophysical datasets, the framework’s performance in different geological settings requires validation. This interpretable, uncertainty-aware approach enhances trustworthiness of predictions in spatially heterogeneous, data-sparse geoscientific problems.

SummaryLower crustal high-velocity bodies (LCHBs) are key indicators of deep magmatic addition and lithospheric modification at rifted continental margins. Integrating 3D gravity modeling with regional geophysical and geological constraints, we identify a prominent LCHB beneath the Xihu Sag of the East China Sea (ECS) shelf basin. This body is NNE–SSW elongated, ~5–7 km thick, and spatially coincides with major depocenters and fault systems. We propose a two-stage mafic emplacement model linking its formation to the tectonic transition from fore-arc compression to back-arc extension. During the early–mid Cretaceous, compressional subduction of the Paleo-Pacific Plate facilitated arc-related underplating and accumulation of mafic material in the lower crust. In the early Cenozoic, slab rollback and asthenospheric upwelling during back-arc extension renewed melt supply, further thickening the lower crust. The absence of surface volcanism indicates that magmas were largely trapped and crystallized at depth, forming dense mafic cumulates. Present-day low shallow-mantle temperatures and high densities beneath the Xihu Sag suggest that preservation of these cumulates was sustained not solely by mantle thermal conditions, but also by prolonged subsidence, sedimentary insulation, and inherited compressional structures. These results underscore the need to integrate tectonic, thermal, and structural factors to fully understand deep magmatic processes in marginal basins.

SummaryWe develop a novel comprehensive theoretical framework, the Biot-patchy-spherical-squirt (BIPSSQ) model, for wave propagation in partially saturated dual-porosity media. This model simultaneously incorporates three key fluid flow mechanisms: macroscopic flow (Biot flow or global flow), mesoscopic flow, and microscopic flow (three-dimensional spherical squirt flow). The constitutive relations and fluid pressure expressions for the BIPSSQ model are first derived and then the governing wave equations are established using a Lagrangian approach based on the system’s kinetic energy, potential energy, and dissipation functions. Through plane wave analysis, we obtain the phase velocity and attenuation of the fast P-wave. Numerical examples demonstrate that the BIPSSQ model predicts multiple dispersion transition bands and corresponding attenuation peaks, attributed to two squirt flows and two Biot global flows from two immiscible fluids. Furthermore, the presence of squirt flow significantly suppresses the mesoscopic patchy saturation effect, leading to the disappear of mesoscopic dispersion transition band and attenuation peak. The influences of permeability, saturation, porosity, squirt-flow length and inclusion radius on velocity dispersion and attenuation are also analyzed. Finally, excellent agreement between theoretical predictions and experimental measurements from an Aksu outcrop rock sample (800 kHz), a gas-water-saturated Estaillades limestone (1 kHz), and an oil-brine-saturated Vosgian sandstone (350 kHz), validates the applicability and effectiveness of the BIPSSQ model. Moreover, the BIPSSQ model can degenerate to other theories (i.e. Biot, BISSQ, BR) under certain conditions. Our proposed model provides a unified and robust tool for interpreting wave propagation phenomena in complex, partially saturated reservoir rocks.

SummaryThe maximum possible earthquake magnitude (${{M}_{MAX}}$) is a consequential parameter that is difficult to quantify. In this paper, order statistics concepts are adapted to infer ${{M}_{MAX}}$ from an earthquake catalogue. Examining jumps in the ordered sequence of largest events significantly improves inferences of ${{M}_{MAX}}$ truncation; I continue this improvement by considering deeper metrics (i.e., jumps in the second, third, fourth, … largest events). I begin by providing a theoretical foundation for these deeper metrics, while highlighting special cases. Synthetic tests are performed to quantify the improvements gained. While the largest information gains arise from the largest event sequence, appreciable gains are found to depths of ten. This approach is also validated on real-data cases, such as Groningen and FORGE, demonstrating their utility. Overall, this approach will contribute to better understanding earthquake hazards and discerning the physical processes that allow earthquakes to grow large.

SummaryThe Ojos del Salado Volcano, the highest active volcano in the world, is located at the southern end of the Puna plateau in central Chile. Here, the subduction angle of the down-going Nazca plate shallows, causing volcanism to move inland marking the southern end of the Central Andean Volcanic Zone (CAVZ). Little is known about the current volcanic activity at this southern edge or the dominant crustal stresses at these volcanic centres. In this study, we use a temporary network of 29 geophones to record local seismicity at the Ojos del Salado volcano. The type of seismic event, number of events per day, location, and magnitudes of events all provide insight into the structure, material properties, and activity level of the volcano. Between February 6th and 28th 2024, this network recorded 93 events with local magnitudes larger than 0, the largest having local magnitude 2.8. The events formed two main clusters, one on the western flank of Ojos del Salado itself near the summit, and a smaller cluster to the north. Most events in the northern cluster occurred within a 35 minute seismic swarm on February 8th. Twenty-one fault plane solutions were determined for events within the network. Six of these occurred during the northern swarm and showed steep oblique faulting and fifteen in the summit cluster, which mainly show normal faulting with strikes comparable to E-W oriented mapped faults in the area. Fault plane solutions at both clusters indicate a north-south extensional stress state. This agrees with the regional stress axes of the southern Puna plateau found in other studies, suggesting that the local crustal stresses at the Ojos del Salado volcano mainly follow the regional stresses with some variation in fault planes near the summit and in the northern swarm that could be due to locally high magmatic or geothermal fluid stresses. Heavy rain in the days preceding the northern swarm may have increased the amount of fluid available, potentially inducing the swarm on February 8th. No seismicity was observed near the Laguna Verde, or the two smaller volcanoes within the network: the Barrancas Blancas and Mulas Muertas. Ojos del Salado is therefore the main source of seismic activity and likely heat source within the study area. The level of seismicity and the occurrence of a seismic swarm to the north and five small seismic swarms near the summit suggest that there is still volcanic activity at Ojos del Salado and it could benefit from monitoring.

SummaryThe eastwards extrusion of Tibetan Plateau (TP) materials has led to complex tectonic deformation and frequent seismicity in the western Sichuan Basin (SCB). To elucidate crustal deformation mechanisms and seismogenic structures, we inverted broadband (3–60 s) Rayleigh wave dispersion curves using ambient noise tomography from 448 stations and constructed a 3-D S-wave velocity (Vs) structure for the upper to middle crust beneath the SCB and adjacent regions. Our model revealed a thick, low-velocity sedimentary layer within the SCB that extends to 15 km depth along its northwestern margin, likely resulting from the accumulation of eroded materials from surrounding orogenic belts. The three-dimensional velocity model resolved sedimentary cover thicknesses ranging from 6 to 13 km within the basin and yielded average Vs values of 3.08, 3.17, and 3.25 km/s for Mesozoic, Palaeozoic, and Proterozoic strata, respectively, thereby calibrating the basement burial depths of major geological units in the sedimentary layers. Notably, this study identified deep low-velocity anomalies beneath the Dayi seismic gap (DSG) and segmented velocity structures along the Kangding–Shimian section, providing crucial deep structural constraints for evaluating the seismogenic environment and future earthquake hazards of major seismic gaps in the Sichuan–Yunnan region. The velocity structure clearly delineates the formation and evolutionary characteristics of multiple foreland basin development episodes since the Late Triassic, offering important constraints for understanding the deep structure of the SCB, assessing seismic hazard risks, and guiding petroleum resource exploration.

SummaryHigh-resolution P-wave velocity tomography of the Japan subduction zone down to 700 km depth is determined by conducting a joint inversion of arrival-time data of local earthquakes and teleseismic events, which were recorded at land-based Hi-net seismic stations and seafloor S-net stations. Our inversion results show the high‐velocity subducting Pacific slab and low‐velocity zones in the mantle wedge beneath active arc volcanoes. Subslab low-velocity anomalies (SLVAs) are revealed in the mantle below the Pacific slab, which may reflect hot and wet mantle upwelling derived from return flow associated with the slab deep subduction. The SLVAs at depths of ~150-260 km exhibit a bimodal distribution, where interplate slow earthquakes occur. There is a SLVA gap below the mainshock hypocenter and rupture zone of the great 2011 Tohoku-oki earthquake (Mw 9.0). The SLVAs may influence the megathrust segmentation by their buoyancy, heat, and melt, and so affect the generation of megathrust and intraslab earthquakes. These results shed new light on the structural heterogeneity and mantle dynamics of the Japan subduction zone.

SummaryThe cycle-skipping problem that plagues full waveform inversion (FWI) can be at least partially mitigated if low frequencies (which encode the kinematics of wave propagation in seismic data) are recorded. However, seismic sources and receivers are band-limited, so seismic data doesn’t generally include signals down to 0 Hz. To improve our ability to solve the seismic inverse problem, one can synthesize this missing low-frequency (LF) content from the recorded high-frequency (HF) data using machine learning (ML) models. Deep learning models such as convolutional neural networks (CNNs) demonstrate impressive ability to perform low frequency extrapolation. However, such models require powerful hardware (GPU machines) and careful training. We assess the extrapolation capabilities of three different ML models that do not require GPU machines, namely, random forest, Gaussian process regression, and gradient boosting, on both synthetic and real data. Experimental results on two synthetic datasets (generated from a low velocity lens embedded in a homogeneous medium, and the Marmousi model) demonstrate that FWI applied to the extrapolated data consistently improves inversion accuracy relative to FWI applied to the original datasets that do not contain low frequencies. Application of low-frequency extrapolation to real data from the Northwest Shelf of Australia demonstrates that tree-based ML models such as gradient boosting can outperform CNNs in terms of both accuracy and computational cost on non-GPU architectures.