In Geological Carbon Sequestration (GCS), reactive transport modeling of CO2-water-rock interactions is computationally prohibitive due to stringent requirements for small time steps and high-resolution grids in traditional numerical solvers. This study develops a novel coupled prediction model (CPM) integrating computational fluid dynamics (CFD) with deep learning to overcome this bottleneck. The method employs a Convolutional Long Short-Term Memory (ConvLSTM) neural network trained on high-fidelity data points generated by OpenFOAM in a real carbonate rock representative elementary volume, capturing a wide range of dissolution patterns. The trained model achieves exceptional short-term prediction accuracy, with normalized root mean square errors of merely 0.62% for concentration and 1.24% for porosity fields, visually indistinguishable from CFD results. To ensure long-term stability, a hybrid strategy of "four steps neural prediction followed by one step CFD correction" is proposed, effectively constraining error accumulation over 40000 s simulations. While accurately reproducing major morphological evolution trends, minor smoothing occurs at complex fluid-rock interfaces. Quantitative analysis demonstrates consistent predictive performance across a range of injection rates within the investigated carbonate REV. Crucially, CPM achieves a 9.8× speedup over conventional CFD under identical computational resources, maintaining physical consistency through periodic physics-based corrections. This work provides a framework for efficient reactive transport prediction, with direct implications for geological carbon sequestration monitoring and reservoir engineering.
Phased array-based look-ahead of drill bits technology serves as a common method in logging-while-drilling (LWD). It utilizes an array of acoustic sources to excite synthetic acoustic wavefields for detecting geological anomalies. Therefore, understanding the spatial–temporal evolution of the multi-source synthesized acoustic wavefield in geological structures is crucial. Previous numerical strategies mainly depend on finite element methods with refined meshes that are computationally expensive. Recently, physics-informed neural networks (PINNs) have attracted considerable attention as a novel paradigm for solving partial differential equations (PDEs). However, PINNs face over-fitting and point-wise weight imbalance challenges when learning multi-source complex wavefields, potentially limiting phased array-based acoustic wavefield reconstruction. This work presents a soft residual attention-based physics-informed neural network (SRA-PINN) to address these problems and reconstruct the phased array-based synthetic wavefield using sparse data. Firstly, the acoustic wave equation is embedded into the loss function as a penalty term to supervise the training process of the deep neural network to characterize the evolutionary process of the synthetic acoustic wavefield. Furthermore, the proposed soft residual attention mechanism assigns point-wise adaptive weights based on the point-wise PDEs residual, thereby enhancing reconstruction quality. Finally, numerical results demonstrate that the proposed SRA-PINN-based method is not only more accurate than the standard PINN but also requires less training data than purely data-driven methods. It efficiently reconstructs multi-source overlapping acoustic wavefields with limited data and without meshes.
The emergence of AI-driven Earth observation systems promises to broaden access to petabyte-scale geospatial data beyond domain specialists. However, translating this vision into operational scientific infrastructure requires addressing fundamental challenges in data virtualization, code transparency, and domain-specific reasoning. We present Hydrology Copilot, a cloud-native AI framework for natural-language-driven analysis of Earth observation data. To demonstrate operational capabilities at scale, we implement the system using NASA's North American Land Data Assimilation System version 3 (NLDAS-3), which provides surface meteorological forcing and land-surface model output across North and Central America at 1-km resolution, from which drought diagnostics are derived.The system integrates five core contributions: (1) scalable data virtualization using Kerchunk-based cloud-optimized access, achieving a 1.5 to 4.6 times improvement in I/O latency across benchmark queries spanning regional single-day extractions (4.6 times speedup) to continental monthly aggregations (1.5 times speedup); (2) transparent code generation through Microsoft Azure AI Foundry agents that expose executable Python workflows for scientific verification; (3) persistent conversational memory enabling multi-turn analytical discourse across sessions; (4) intelligent query validation that enforces dataset boundaries and resolves ambiguous requests before execution; and (5) a multi-agent architecture coordinating query parsing, code generation, and visualization. We evaluate the system through drought-monitoring workflows, demonstrating reliable code generation, accurate results validated against reference computations and the operational U.S. Drought Monitor, and efficient operation across increasingly complex tasks.By bridging natural-language interfaces with rigorous hydrological analysis, Hydrology Copilot advances beyond proof-of-concept demonstrations to provide a deployable framework for operational Earth science applications.
Full waveform inversion (FWI) using seismic data holds great promise for resolving fine-scale subsurface structures and has emerged as a powerful tool for monitoring time-lapse CO2 plume dynamics, particularly in vertical seismic profiling (VSP) applications. However, challenges persist in field deployments, where seismic data are typically band-limited, noisy, sparsely sampled, and often incomplete. Furthermore, variations in near-surface conditions between surveys—such as changes in source and receiver coupling and positioning under differing environmental conditions—introduce substantial inconsistencies in the recorded data. These factors render time-lapse FWI a highly underdetermined and ill-posed inverse problem, where multiple models can explain the data equally well, posing a serious threat to reliable CO2 plume monitoring and decision-making. To address this, we propose a deep neural network representation multiphysics joint inversion framework. In this approach, geophysical data from different domains—such as gravity and resistivity—serve as natural constraints for seismic inversion, since they represent complementary physical properties of the same subsurface region. The proposed framework differs from conventional cross-gradient joint inversion in that the coupling among elastic, density, and electrical parameters is achieved through an implicit parameter-mapping mechanism in a shared neural representation, rather than through manually imposed structural similarity constraints. The latent features generated with the networks are projected to respective physical domains with appropriate grid resolutions for forward modeling. In this paper, we will focus on the joint inversion of seismic, gravity, and resistivity data, and data in other geophysical domains can also be implemented directly. Numerical experiments demonstrate that the proposed method can effectively recover the spatial migration of injected CO2, while apparently accelerating the convergence of density and resistivity models. Vp, and Vs model can also be improved due to the shared usage of the weights in the network. Under non-ideal conditions, characterized by high noise levels and survey inconsistencies between time-lapse datasets, our approach yields more robust and interpretable inversion results, highlighting its potential for reliable long-term CO2 monitoring.
A fully numerical framework is presented for gravity inversion to estimate the three-dimensional volumetric density distribution and basement structure of sedimentary basins. In the spatial domain, gravity anomalies of prismatic bodies are computed using Nagy’s analytical formulation, while in the frequency domain a modified Parker’s modal approach combined with the Fast Fourier Transform using Gaussian quadrature is employed to handle complex, irregular basin geometries and heterogeneous density distributions efficiently. Forward responses are perturbed with Gaussian noise to mimic realistic measurements and are inverted using a cubic B-spline representation of the basement, defined on a coarse control grid. The inversion minimises a regularised objective function that combines variance-normalised gravity misfit and a two-dimensional Laplacian smoothness constraint on the B-spline control points. A global–local optimisation strategy is adopted, where a differential evolution algorithm first explores the full parameter space and the best model is subsequently refined using a Limited-memory Broyden–Fletcher–Goldfarb–Shanno with Bound constraints algorithm. The applicability and robustness of the approach are demonstrated on synthetic basin models and further validated through real-case modelling of the Chintalpudi sub-basin, India, showing its precision, computational efficiency, and adaptability for three-dimensional subsurface geological investigations.
Continuation is an essential method in potential field theory. Compared with Upward continuation (UC), Downward continuation (DC) is unstable. Traditionally, solutions are obtained through optimization algorithms and regularization conditions, however, the difficulties lie in selecting the regularization parameter. Neural networks offer a novel solution to this problem, but their interpretation is limited. Based on these, in the comparative simulation of UC and DC, while considering the universality suitable for different height parameters, we constructed a variable-height and stable continuation method (ContiNet) based on the basic multi-layer perceptron (MLP). For UC, based on the structural similarity between MLP and UC, we simulated the discrete UC. And then we constructed the UC neural network (UpContiNet), which enabled the effective simulation of the theoretical UC. For the DC, the concept of weighted regularization based on continuation height and neural network was integrated into UpContiNet. Thereby, we constructed the DC neural network (DnContiNet), which achieved stable theoretical DC. Compared with other methods, numerical simulations and field surveys all indicate that DnContiNet performs well in the DC problem. The intuitive diagrams and error indicators have all improved. The network construction idea and model have reference value for research and application in the field of geophysics.
Automatic shoreline extraction from remote sensing imagery remains challenging in complex coastal environments, where tidal-flat transition zones, surf-zone foam, turbid waters, and artificial shorelines often cause boundary serration, local discontinuities, and positional drift. Using Sentinel-2 Level-2A imagery, the self-constructed Vietnam dataset was divided into 1571 training and 393 validation image–label patches, each with a size of 256 × 256 pixels. This study proposes a boundary-enhanced convolutional neural network–Transformer model with a bi-directional cascade network-style boundary branch, termed CT-BDCN, for binary sea–land segmentation and fine shoreline localization. The proposed model integrates a multi-scale boundary branch, gated boundary enhancement, and boundary residual refinement to strengthen shoreline-neighborhood feature representation. A shoreline-neighborhood evaluation scheme was further constructed, and a narrow-band edge intersection-over-union metric, termed shoreline-neighborhood edge intersection over union (IoU_edge), was used to quantify prediction consistency near the shoreline. CT-BDCN contains 1.1403 million parameters and requires 23.0183 giga floating-point operations for an input size of 1 × 3 × 256 × 256.Experimental results show that, although the improvement in region-level segmentation accuracy is relatively limited, the proposed method substantially enhances shoreline boundary quality. Compared with the baseline convolutional neural network–Transformer model, CT-BDCN improves intersection over union (IoU) and IoU_edge by approximately 1.31% and 16.79%, respectively. Compared with U-Net, DeepLabV3+, and SegFormer, IoU_edge is improved by approximately 9.29%, 26.97%, and 27.22%, respectively. These results demonstrate that explicit multi-scale boundary modeling effectively improves shoreline continuity and localization accuracy in complex coastal environments, providing technical support for large-scale and fine-grained shoreline extraction.
Large-field-of-view (FOV) FE-SEM mosaics provide high-resolution two-dimensional images of shale pore systems, yet quantitative pore statistics derived from such images are strongly influenced by acquisition resolution and segmentation consistency. In this study, a reproducible workflow integrating multi-resolution FOV FE-SEM imaging, manual annotation, deep learning–based semantic segmentation, and extraction of image-derived pore statistics is proposed to systematically quantify resolution sensitivity under a constant field of view (400 × 400 μm). Four convolutional neural network models were trained and evaluated against manually labeled datasets, achieving good agreement with manual segmentation, enabling consistent pore-mask generation across resolutions without per-image threshold tuning. Using FE-SEM mosaics acquired at pixel sizes ranging from 500 to 4 nm/pixel, we quantify the resolution dependence of image-derived two-dimensional surface porosity (2D porosity), pore and fracture counts, pore-size distributions (PSD), and box-counting fractal dimension (FD). Results show that detectable pore and fracture counts increase markedly with finer pixel size, whereas area-based 2D porosity shows comparatively weaker resolution dependence once the pixel size is finer than approximately 16 nm/pixel. PSDs exhibit resolution-dependent lower-size truncation controlled by the image detection limit, micro-pores and micro-fractures dominate numerically, whereas larger features contribute more strongly to the segmented 2D area. FD shows a non-monotonic relationship with resolution, reflecting the combined effects of boundary simplification at coarse resolution and noise amplification at fine resolution. Comparison with low-pressure CO2/N2 adsorption demonstrates that FOV FE-SEM and adsorption measurements provide complementary pore-size information. These findings provide a resolution-aware basis for interpreting FE-SEM-derived 2D pore statistics and for selecting imaging parameters under acquisition and computational constraints.
The collection of airborne geophysical data, ranging from ice sheet thickness measurements to magnetic anomaly detection for critical mineral exploration, continues to be of increasing importance. Interpolation of airborne geophysical data is essential, due to wide gaps between survey flight lines. Traditional interpolations are deterministic, meaning they produce a single map of predicted values between the data locations. This substantially limits the utility of the interpolation, as there is no acknowledgment of the fact that there is no true knowledge (e.g., no data was sampled) at these locations. These deterministic interpolation methods are thus limiting for downstream actions, hindering decision making and obscuring reality, which is uncertain. Geostatistical tools, such as sequential Gaussian simulation, have been deployed to quantify spatial uncertainty via stochastic rastering, but suffer from interpolation artifacts due to the geometric anisotropy of airborne survey data acquisition (e.g., dense along-track, sparse across-track). Existing methods that aim to reduce interpolation artifacts have been developed for deterministic interpolation schemes, but have yet to be carefully considered for stochastic schemes, leaving a clear gap in the interpolation toolbox: a method which properly quantifies spatial uncertainty while simultaneously suppressing unphysical interpolation artifacts. To address this gap, we introduce a new multigrid-based simulation method. This multigrid simulation is a stochastic geostatistical approach that mitigates interpolation artifacts while generating an ensemble of many realizations by treating the domain as a random field. Multigrid simulation is able to handle interpolation across large-scale domains, including non-stationary domains. We demonstrate its effectiveness on two datasets, one synthetic and one from a real field survey.
Accurate segmentation of paleontological fossil images is of great scientific significance for morphological studies, paleoinformatics, and downstream analysis applications. However, fossil images generally exhibit high similarity between the subject and background, blurred edges, and loss of fine details, which challenge mainstream neural networks to achieve reliable segmentation. To address these issues, we construct the first fossil segmentation dataset, FOSSIL3K, containing 3289 high-quality pixel-level annotated images covering multiple typical fossil categories. Based on this dataset, we propose a fossil segmentation network, FRNet, which incorporates a Frequency Fusion Module (FFM), Boundary-Aware Module (BAM), and Foreground Refinement Module (FRM) to enhance object localization and edge detail recovery. FRNet first leverages fused frequency and high-level features to accurately locate the main object regions; then, by integrating foreground, background, and edge information, it progressively refines the target regions to restore structural and micro-level details. Experimental results on FOSSIL3K demonstrate that FRNet significantly outperforms state-of-the-art Salient Object Detection (SOD) and Camouflaged Object Detection (COD) methods, and also achieves leading performance on four COD benchmark datasets. These results validate the effectiveness of FRNet for fossil image segmentation and demonstrate the feasibility of adapting COD techniques to domain-specific tasks.
Accurate delineation of both pore space and rock matrix is critical for petrophysical investigations in sandstone reservoirs. Conventional methods struggle with the complexity of micro-CT data, while deep learning approaches require costly pixel-level annotations. Few-annotation segmentation mitigates annotation demands; however, existing approaches are still susceptible to overfitting and show limited generalization. To address this challenge, RockAdaptive-FewSeg, a novel few-annotation segmentation framework that integrates a Residual Attention U-Net architecture with adaptive watershed-based cropping augmentation, optimized inference strategies, and an Expert-in-the-Loop correction mechanism, is proposed. A five-stage progressive evaluation conducted on the benchmark Ottawa sand CT dataset demonstrates the effectiveness of the proposed framework under an extremely sparse supervision regime, where pixel-level annotations are provided for only one or two images in the entire training set. The results demonstrate that, on the Ottawa sand evaluation dataset, the segmentation accuracy improves consistently across stages, achieving Dice of 0.995, Intersection over Union (IoU) of 0.991, and a 79.7% reduction in average surface distance (ASD), which highlights the superior boundary preservation enabled by the progressive optimization strategy. Moreover, the digital rock models derived from the final segmentation yield highly accurate petrophysical predictions, with a mean absolute percentage error (MAPE) below 4%. The framework provides a practical and annotation-efficient solution for generating high-quality digital core data on the Ottawa sand benchmark, demonstrating potential for reliable petrophysical analysis in geo-energy research.
Automatic mineral image identification plays a crucial role in geological exploration and related applications. Although multimodal data, such as physical attributes and spectral information, can enhance recognition accuracy, field environments often suffer from missing or corrupted modalities, leading to reduced system robustness. This paper integrates Scale and Shift Feature adaptation (SSF) and a Dynamic Multimodal Information Bottleneck (DMIB) mechanism to improve model stability and generalization under modality-missing or erroneous conditions. The framework is parameter-efficient, requiring only about a 13.8% increase in parameters compared to the baseline while enabling adaptive recognition across various modality combinations. We evaluate the proposed method on a self-constructed dataset containing 36 mineral classes and more than 110,000 images. The SSF-DMIB achieves 93.7% accuracy under complete-modality conditions, which is 1.3% higher than the baseline, and maintains 71.4% accuracy under extreme scenarios with 100% modality corruption, surpassing the baseline by 28.3%. These results demonstrate that SSF-DMIB exhibits strong robustness and practicality in complex field environments and provides a promising solution for real-world multimodal mineral recognition.
Bioturbation is critical for reconstructing sedimentary paleoenvironments, predicting stratigraphic distribution, and elucidating the mechanisms by which organisms modify reservoir quality. Conventional methods for identifying bioturbation rely mainly on manual visual interpretation. These approaches are often limited by significant subjectivity, low efficiency, and challenges in quantifying subtle morphological variations. To overcome these limitations and achieve a fundamental shift from semi-quantitative grading to continuous quantification, we developed the U2-MFALNet (U-squared Multi-scale FPN-ASPP-LSKA Network). This architecture is specifically engineered to precisely delineate biogenic structures at the pixel level, thereby enabling the automated and accurate calculation of bioturbation intensity. Specifically, the network incorporates a Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale context, while integrating Large Selective Kernel Attention (LSKA) to extract irregular biogenic structures, complemented by a Surface-Aware Loss function to refine boundary segmentation. In this study, we constructed a dataset comprising 3028 high-resolution images derived from Ordovician carbonate cores in 16 wells of the Tahe Oilfield, Tarim Basin. For model training and validation, we utilized refined pixel-level labels that were rigorously calibrated by comprehensive geological, logging, and mud logging data. Experimental results demonstrate that the proposed model achieves a mean Intersection over Union (mIoU) of 90.85%, yielding a 3.66% improvement over the baseline U2-Net, and significantly outperforming mainstream baselines such as U-Net, DeepLabV3, and SegNeXt. Furthermore, quantitative assessments confirm the model's high precision in calculating the Bioturbation Index (BI), validating its reliability for objective geological analysis. This research offers a robust solution for the automated, continuous quantification of bioturbation, supporting the advancement of intelligent and quantitative sedimentology.
Spectral analysis based on the fast Fourier transform (FFT) provides a powerful framework for quantifying land surface variability across spatial scales, but its application to digital elevation models (DEMs) is often limited by spectral leakage due to boundary discontinuities and by land surface non-stationarity. This study presents fft-tools, a modular, computationally efficient toolkit for spatially explicit spectral analysis of DEMs. The approach combines block-based localized Fourier decomposition with an outer tapering strategy based on thin-plate spline extrapolation combined with residual reconstruction, which suppresses boundary-induced artifacts while preserving the full interior signal. The framework supports three core workflows: (1) spectral characterization of land surface structure using polar transformation and radial and angular analysis of the power spectral density (PSD), (2) scale-specific land surface reconstruction via frequency-domain filtering and inverse FFT, and (3) quantitative comparison of DEMs using PSD ratios and spectral coherence. Application to a geologically structured mountainous area demonstrates the ability to resolve anisotropic landform organization, distinguish lithologically controlled surface textures, and isolate land surface components across a wide range of spatial scales, from meter-scale roughness to large-scale morphology. These results highlight the potential of the approach for quantitative, scale-dependent analysis of the land surface as a physically meaningful signal, supporting the development of physical geomorphometry.
Geoscientists interpret thin section images through the lens of their own training and experience, assisted by computational systems only to perform routine descriptive activities. This study introduces an interactive computational framework to quantify and analyze how scientists at different levels (undergraduate students, graduate students, mid-career and senior researchers) perceive the same petrographic thin sections. By logging parameter selections and segmentation choices in a controlled environment, we reveal systematic differences in perception, focusing on porosity, connectivity, and pore-type classification. Our results highlight the importance of computational tools in bridging subjective interpretations, reducing uncertainty, and creating reproducible workflows for petrographic analysis.
Q-compensated prestack depth migration (QPSDM) is an effective tool for achieving 3D high-resolution imaging of subsurface structures. Although optimized strategies like the optimal migration aperture and multiprocessing GPU/CPU heterogeneous parallel schemes have been adapted from Q-compensated prestack time migration (QPSTM) to reduce computational cost, the efficiency of GPU-QPSDM remains a challenge for large-scale seismic data. These challenges include imbalanced data distribution, inconsistent GPU performance, and resource contention caused by concurrent tasks. To address these issues, we propose a new multithreading parallel computing scheme. Instead of using multiple processes, our approach launches multiple parallel reading and computing threads, effectively overlapping data I/O with computation. This scheme effectively reduces I/O pressure, minimizes redundant resource usage, and mitigates delays from uneven GPU performance, ensuring balanced task execution. Furthermore, A series of scaling numerical tests confirmed that our multithreading scheme consistently outperforms the multiprocessing approach, demonstrating strong scalability with increasing data size. Our scheme is available to both QPSTM and QPSDM algorithms. A numerical example based on large-scale 3D seismic data with QPSDM demonstrates significant performance improvements compared to the multiprocessing scheme, which obtained obvious resolution enhancements compared to the conventional prestack depth migration (PSDM), highlighting its potential for efficient 3D high-resolution seismic imaging in industrial applications.
Deep learning-based fault interpretation has recently emerged as a dominant approach, offering significant advancements in automation and accuracy. However, its effectiveness is often constrained by the reliance on large amounts of expert-labeled data. Moreover, the sparse and elongated nature of faults makes them difficult to identify accurately, especially when only limited annotations are available. To overcome these challenges, we propose a novel Uncertainty-driven Region-based Active Learning (URAL) workflow. URAL employs an uncertainty-based sampling strategy to select informative samples and uses coordinate-based K-means clustering to identify sample-dense regions as annotation areas, directing expert attention to fault-rich regions. Additionally, we introduce an automatic cropping strategy along fault extension directions to increase the fault-to-background ratio in training patches, mitigating the severe imbalance between faults and non-faults. Applied to Thebe, a real large-scale, well-labeled seismic dataset, URAL achieves performance comparable to fully supervised learning using only 3.7% of the labeled data. Moreover, when compared to classical AL methods, URAL achieves similar accuracy with only 21% of the labeled pixels, further demonstrating its ability to minimize labeling effort while maintaining high performance. The proposed method offers a scalable solution for large-scale seismic fault interpretation.
When spatial data are abundant, complex geographical phenomena can be effectively modeled using multiple-point statistics with observed data as training images. However, two major challenges remain: the stationarity assumption of spatial processes is frequently impractical, and resampling from observed data fails to ensure the continuity and completeness of variable values. To address these limitations, this study developed a multiple-point statistics simulation method based on Yang Chizhong filtering (YangMPS). We propose that a spatial variable comprises regular (non-stationary trend) and irregular components, with only the latter being the target for the multiple-point simulation. When the trend is considered as a linear variation within a local region, we mathematically proved that Yang Chizhong filtering can provide an approximately optimal unbiased estimate of the trend. Therefore, the regular component was first constructed using generalized Yang Chizhong filtering and interpolation method. We identified multiple-point patterns from the irregular component via density peak clustering and derived the cumulative distribution function of each cluster via kernel density estimation. Using the cumulative distribution functions of the clusters, a Monte Carlo simulation method was developed to reconstruct the irregular component, which was fused with the trend for the final result. Experiments on simulated and geochemical datasets showed that YangMPS achieved significantly higher accuracy in modeling non-stationary processes than the existing direct sampling and detrending methods. The attribute values of the simulation results obtained by YangMPS can be continuously taken from the cumulative distribution functions of the clusters.
