Geophysical Prospecting - Volume 73, Issue 6, 2025

Digital rock techniques are increasingly important in petroleum exploration and petrophysics. Digital rocks are typically acquired via scanning or imaging techniques, but the resulting images may lack clear, detailed information due to resolution limitations. Super‐resolution reconstruction using deep learning offers new possibilities for digital rock technology development. In current research on super‐resolution reconstruction of digital rock images, most networks employ attentional mechanisms in a single dimension, ignoring more comprehensive interactions from both spatial and channel dimensions.

To address the above problems, we propose a bi‐dimensional large kernel attention network for super‐resolution reconstruction of digital rock images. The network consists of three components: a bi‐dimensional large kernel building block, a contrast channel attention block and an enhanced spatial attention block. In addition, the traditional method of stacking network modules to build the network leads to an increase in computation and network size, so we adopt Transformer's MetaFormer architecture, which integrates multivariate feature extraction to improve the efficiency of the network. In the process of feature information circulation, we effectively prevent shallow feature loss by two efficient attention modules working at different network depth positions. Extensive experiments on Sandstone2D and Carbonate2D rock datasets show that our proposed model significantly outperforms existing image super‐resolution networks.

The transformation of elastic impedance (EI) from partial‐angle‐stacked seismic data is a crucial technique in the domains of reservoir modelling. Conventionally, EI inversion is performed on a per‐angle basis, leading to significant discrepancies in EI values across different angles, which may not accurately represent actual conditions. When the signal‐to‐noise ratio (SNR) of seismic data is low, the inverted EI tends to be unstable, resulting in poor‐quality inversion outcomes. This research proposes a novel method that allows for enabling the derivation of EI for various angles simultaneously inverted from multiple partial angle‐stack seismic datasets in one process. The aim of simultaneous inversion is to potentially ensure consistent EI results. To obtain this aim, we utilize an advanced regularization method called the Gramian constraint. Consequently, the objective function for the simultaneous inversion of multiple EIs is developed. Results from both synthetic and field data demonstrate improved stability in EI inversion, especially for the case of low SNR.

Time‐lapse seismic signals and their relation to variations in reservoir pore pressure and fluid saturations are, in general, well understood. Occasionally time‐lapse (4D) seismic data do present some intriguing anomalies that cannot be properly explained by our general well stablished expectations, forcing us to consider less conventional hypotheses. We present one such case, occurring in a weakly cemented sandstone reservoir in the North Sea. This reservoir presents a few 4D seismic softening signals occurring as a response to pore pressure decrease, where no saturation changes are expected. With a detailed multidisciplinary analysis, we assess all possible explanations for this type of signal and show that conventional explanations in terms of fluid saturation changes and/or elastic stress variations fail to explain the full characteristic of the observed anomalies. As an alternative hypothesis, we propose the possibility of pore collapse, an inelastic rock damage process, as an unconventional explanation to the observed anomalies. We show that this hypothesis is the only one that explains all the characteristics of the observed anomalies in terms of their lateral and vertical extents, as well as their magnitude and temporal evolution behaviour. We then conduct a theoretical modelling feasibility study to estimate the critical pressure for initiation of rock damage and estimate the amount of rock damage needed to produce the observed 4D seismic signals. This feasibility study suggests that the reservoir effective pressure achieved during field production is likely not enough to crush grains and cause reservoir compaction. However, they may be enough to cause cement and weak grain cracking, which we estimate through rock physics modelling to be a sufficient mechanism for producing the observed softening anomalies. This makes weakly cemented sandstones more prone to this type of counterintuitive signal, as cement damage occurs at lower effective pressures, more commonly achieved during reservoir production. We also highlight important considerations regarding plans of CO2 storage into depleted reservoirs, as the possibility of rock damage during production would complicate the monitorability of the injected CO2 plume.

Radiomagnetotellurics (RMTs) is an efficient frequency‐domain electromagnetic technique for mapping subsurface electrical resistivity, particularly suited for near‐surface investigations. This method utilizes commonly available civil and military radio transmitters, broadcasting between 10 kHz and 1 MHz, as sources to measure electric and magnetic field responses at the surface. Modern RMT receiver systems comprise five components (two electrical antennas and three magnetic coils), allowing for the estimation of the full impedance tensor and the tipper transfer function for the vertical magnetic field. In this study, RMT data were acquired to investigate the shallow structure of the Himalayan Frontal Thrust (HFT) fault in the Sub‐Himalayan region around Uttarakhand, India. Data were collected at 312 stations along eight profiles over an area of roughly 500 m × 70 m. The dense station distribution enables a 3D inversion of the dataset in the extended frequency range of up to 1 MHz. The observed data were processed using scalar as well as tensor estimations to obtain full impedances and tipper transfer function. We integrated scalar‐estimated data from zones with an approximately 2D conductivity distribution in the full‐tensor dataset. This approach ensured robust 3D modelling during the initial RMT inversion performed with the ModEM algorithm. To date, a joint 3D interpretation of RMT full impedance tensor and tipper transfer function has not yet been reported. Furthermore, the near‐surface manifestations of the HFT have not previously been explored by RMT. The derived 3D model from combined scalar, tensor and tipper data reveals a conductivity contrast zone that aligns well with the HFT fault outcrop and complementary geological information. The derived geo‐electrical structure recovers the local sediment thickness and shallow fault inclination.

We compute probabilistic Niobium–Yttrium–Fluorine (NYF) pegmatite prospectivity maps in the Tysfjord region in Northern Norway. NYF pegmatites are generally enriched in rare earth minerals and represent residual melts derived from granitic plutons or melts formed by partial melting of metaigneous rocks. In Tysfjord, however, these pegmatites contain high‐purity quartz, which is the major target commodity of exploration and mining. As the area is geologically underexplored, we employ a data analytics approach for the discovery of new deposits. We carefully lay out our knowledge base and how it impacts the working hypothesis and feature engineering. Self‐organizing maps are employed as an unsupervised and random forest classification as a supervised data analytics algorithm to process and link features derived from airborne magnetic and radiometric maps with sparse pegmatite occurrences available in the form of outcrops and active and abandoned mines. The predictive power of our probabilistic pegmatite prospectivity maps is analysed by means of additional boreholes, which indicates the usefulness of our prospectivity maps for exploration targeting. We recommend employing unsupervised and supervised data analytics approaches in exploration targeting case studies where uncertainty about the predictive power of the available database cannot be ruled out before subjecting the database to data analytics.

Borehole seismic investigations play a major role for high‐resolution imaging of geological structures at depth. The resulting borehole seismic data enable a direct characterisation of the target units as well as their physical properties along the well and in its direct vicinity. Analysing seismic data acquired at different scales within the borehole provides additional notable insights and allows an improved geological and petrophysical interpretation. In our work, we processed zero offset vertical seismic profiling data and full waveform sonic log data as part of a multiscale borehole seismic imaging workflow to better characterise a mineral exploration target at Ludvika Mines (Blötberget mining area, Central Sweden). Data processing mainly comprised wavefield separation and corridor stacking, followed by migration of the full waveform sonic log data using a diffraction stack approach. Additional borehole data, that is, impedance logs and a lithological borehole profile, were used for the integrated interpretation to provide the basis for an assignment of the reflectors to a specific lithological unit. Besides the existing structural models derived from surface seismic investigations, the new images from borehole seismic data reveal the internal structure of the mineralisation at a significantly higher resolution, complement the geophysical characterisation and can be used for a subsequent reliable mineral resource estimate.

Distributed fibre‐optic sensing (DFOS) technologies have emerged as cost‐effective high‐resolution monitoring alternatives over conventional geophysical techniques. However, due to the large volume and noisy nature of the measurements, significant processing is required and expert, fit‐for‐purpose tools must be designed to interpret and utilize DFOS measurements, including temperature and acoustics. Deep learning techniques provide the flexibility and efficiency to process and utilize DFOS measurements to estimate subsurface energy resource properties. We propose a deep learning‐based dual latent space method to process distributed acoustic sensing (DAS) and distributed temperature sensing (DTS) measurements and estimate the injection point location and relative multiphase flow rates along a flow‐loop equipped with a DFOS unit. The dual latent space method is composed of two identical convolutional U‐Net AutoEncoders to compress and reconstruct the DAS and DTS data, respectively. The AutoEncoders are capable of determining an optimal latent representation of the DAS and DTS measurements, which are then combined and trained using one experimental trial and used to estimate the physical flow properties along five different test experimental trials. The predictions are obtained within 7 ms and with over 99.98% similarity and less than absolute error. The method is also shown to be robust to Gaussian noise and can be applied to different multiphase scenarios with a single pre‐training procedure. The proposed method is therefore capable of fast and accurate estimation of physical flow properties at the laboratory scale and can potentially be used for rapid and accurate estimation in different laboratory or field subsurface energy resource applications.

The overburden structures often can distort the responses of the target region in seismic data, especially in land datasets. Ideally, all effects of the overburden and underburden structures should be removed, leaving only the responses of the target region. This can be achieved using the Marchenko method. The Marchenko method is capable of estimating Green's functions between the surface of the Earth and arbitrary locations in the subsurface. These Green's functions can then be used to redatum wavefields to a level in the subsurface. As a result, the Marchenko method enables the isolation of the response of a specific layer or package of layers, free from the influence of the overburden and underburden. In this study, we apply the Marchenko‐based isolation technique to land S‐wave seismic data acquired in the Groningen province, the Netherlands. We apply the technique for combined removal of the overburden and underburden, which leaves the isolated response of the target region, which is selected between 30 and 270 m depth. Our results indicate that this approach enhances the resolution of reflection data. These enhanced reflections can be utilised for imaging and monitoring applications.

CO2 geological sequestration (CGS) is a crucial strategy to mitigate the greenhouse effect. The quantitative correspondence between CO2 saturation and acoustic response serves as the essential basis for monitoring CO2 migration. However, due to dynamic fluid interactions between supercritical CO2 and brine/oil in porous media, acoustic propagation behaviour is extremely complicated, even at the same saturation during drainage and imbibition processes. This study is motivated to evaluate the acoustic characteristics of the above porous stratum containing CO2. To do so, pore fluid parameter models specific to CGS are consolidated and refined, with the consideration of CO2 solubility. Meanwhile, Lo's theory is modified to describe both partial flow and global flow in CO2‐saturated porous media, capturing key mechanisms of patchy distribution and alterations in capillary pressure and relative permeability during drainage and imbibition. By combining these procedures, the wave propagation characteristics within CGS scenarios are systematically analysed. It is shown that CO2 exhibits higher solubility than gases, leading to a distinct two‐stage acoustic response, corresponding to its dissolved and free states. Relative permeability affects both compressional and shear waves, whereas capillary pressure and patchy distribution mainly affect compressional wave propagation. Notably, compressional waves exhibit heightened sensitivity to free CO2 content and fluid flow dynamics, especially at ultrasound frequencies. The modified acoustic propagation theory demonstrates superior performance in characterizing compressional velocities during both drainage and imbibition. These findings highlight the dynamic fluid flow effects in CGS, providing a theoretical framework for analysing acoustic propagation characteristics.

Absolute impedance estimation is crucial for quantitative interpretation of petrophysical parameters such as porosity and lithology, from band‐limited seismic data. The missing low‐frequency part of the conventional seismic data leads to non‐uniqueness in the solution and causes a hindrance to the absolute impedance estimation. This work presents an application of seismic envelope to retrieve absolute acoustic impedance (AI) values directly from band‐limited data in an innovative workflow based on a deep sequential convolutional neural network (DSCNN). Along with the band‐limited data and seismic envelope, we also incorporate the instantaneous phase information (to compensate for the lost phase information in a seismic envelope) as an auxiliary input into the DSCNN model to map the band‐limited data into broadband data and then to retrieve absolute AI values. We have tested the proposed workflow on two synthetic benchmark datasets of Marmousi2 and SEAM 2D subsalt Earth model, as well as one field dataset of the F3 block, the Netherlands. Our results underline that the proposed approach is efficient in recovering the deeper features quite well as compared to the conventional approach, wherein only seismic band‐limited data are used as input. Numerical tests show that the estimated low‐frequency impedance is recovered well with our proposed seismic envelope‐driven approach. Thus, the proposed workflow provides a robust solution for broadband impedance inversion by utilizing only one regression‐based unified deep learning (DL) model. This work primarily highlights the potential of seismic envelope to greatly improve the estimation of low‐frequency components of subsurface impedance model in a DL framework. Such a workflow for absolute impedance inversion from band‐limited seismic will play an important role in reservoir characterization and in quantifying the elastic attributes.

A comprehensive deep learning approach was introduced, encompassing data denoising, inversion imaging and uncertainty analysis. For denoising transient electromagnetic (TEM) data, we utilized a Bidirectional Long Short‐Term Memory (BiLSTM) network. In the data inversion process, a combination of convolutional neural network (CNN) and BiLSTM structures was employed, and their outputs were consolidated using a multi‐head attention mechanism. To ensure robust performance under challenging noise conditions, we implemented a specialized multi‐channel noise training protocol during model optimization. The framework incorporates Monte Carlo (MC) dropout techniques to systematically evaluate prediction reliability throughout the inversion pipeline. This approach has not only been validated on test datasets but has also been successfully applied to the field dataset collected at the Narenbaolige Coalfield in Inner Mongolia, China. The deep learning inversion results obtained from both raw and denoised data exhibit reduced vertical continuity and increased roughness characteristics. In contrast, the Occam's inversion method with smoothness constraints yields results demonstrating superior lateral continuity and vertical smoothness. It is noteworthy that both inversion approaches show consistent interpretations regarding the scale of basalt formations and their contact interfaces with underlying sedimentary layers. Further uncertainty analysis reveals relatively higher uncertainty characteristics in the transition zones between basalt and sedimentary layers, as well as in deeper formations. The elevated uncertainty at interface regions may be attributed to model resolution limitations and inversion ill‐posedness issues, whereas the higher uncertainty in deeper formations is more likely caused by the volumetric effects of electromagnetic field detection and the influence of observational data noise.

Accurate simulation of seismic waves is essential for achieving high‐precision full‐waveform inversion (FWI). Within the Cartesian coordinate system‐based frequency‐domain finite‐difference (FDFD) framework, we propose a one‐direction composition average‐derivative optimal method for the 3D heterogeneous isotropic elastic‐wave equation, referred to as the 45‐point scheme. The results of dispersion analysis and weighted coefficient optimization demonstrate that the 45‐point scheme achieves higher dispersion accuracy than the existing 27‐point average‐derivative scheme. More importantly, by constructing the impedance matrix along the ‘composition’ direction, the bandwidth of the sparse impedance matrix increases only slightly, with nonzero elements compactly distributed in strips. On the basis of the multifrontal massively parallel sparse direct solver (MUMPS) on a supercomputer platform, the 45‐point scheme does not significantly increase computational complexity compared to the 27‐point scheme. To further test the performance of the 45‐point scheme, we provide several numerical experiments, including simple homogeneous and complex SEG/EAGE overthrust models. In comparison with the 27‐point scheme, the 45‐point scheme yields a notable improvement in computational accuracy, particularly for large grid ratios, while imposing only a modest increase in computational cost. These findings thus strongly suggest that the 45‐point scheme holds promise as a viable option for the forward part of frequency‐domain FWI in practical high‐accuracy seismic imaging applications.

The data recorded by the seismometer on the InSight are contaminated by interference signals called ‘glitches’, which have a specific duration and waveform. These glitches emerge very frequently with large amplitude differences and affect the subsequent processing of the data. Traditional methods for glitches removal rely on the threshold and cannot perfectly detect non‐standard and composite glitches. We propose a deep learning based method for glitch detection and removal. A detection model based on the PhaseNet network is developed for three‐component data. The ConvTasNet from the field of speech signal separation is introduced into the noise removal model to separate the glitches from the single‐component data. The advantages of deep learning include the ability to autonomously extract features from the training set without requiring parameter adjustment and the ability to quickly process large amounts of data. The proposed method can detect and suppress non‐standard glitches and provides a novel approach to removing them from Mars exploration records.