Abstract Salt formations often consist of layered evaporite sequences (LES) with mechanically distinct lithologies, which generate complex internal structures during deformation. Interpreting these structures in seismic data is challenging because of poor imaging, discontinuous reflectors, and strong velocity contrasts. While machine learning (ML) has been increasingly applied to salt boundary detection, its potential for resolving intra-salt structures remains underexplored. This study evaluates the performance of two supervised ML tools—fault prediction and multi-horizon interpretation—across three 3D seismic surveys from the Norwegian North Sea, where the Zechstein salt ranges from bedded salt to salt pillows and diapirs. Results show that ML performance is strongly dependent on seismic image quality and the degree of deformation: bedded salt allows accurate tracking of top, base, and intra-salt horizons, whereas pillows and diapirs exhibit degraded reflector continuity and lower interpretation confidence. Fault prediction not only detects faults but highlights intra-salt discontinuities such as rotated stringers. Confidence maps provide uncertainty-aware interpretations that delineate the zones of higher risk. In the context of subsurface hydrogen storage, the ML-derived confidence maps can guide salt cavern placement and reduce geomechanical risk.
The identification of volcanic lithology is crucial for the exploration and development of oil and gas reservoirs. Current methods primarily rely on core and thin-section data, which, despite their accuracy, are often inefficient, costly, and ubjective. To addres thee limitations, conventional logging curvs ss secalibrated with core and thin-section data were sed. A random forest (RF) intelligent idntificaton ueimodel was establshed, and a novl ADASYN-KNN-RF ensemble model was further proposed for volcanic lithology identificatin. The sample numbrs of various lithologies were first balanced using the ADASYN delta oeoversampling algorithm. Subequently, KNN ad RF models were trained independently. An optimal snclassification strategy was then determined by evaluaing the performance of the two models on different tlithologie, and their ults were effectively integrated. When applied to actual data, he proposed ensemble srestmodel achieved a high identification accuracy of 93%. Coparative analysis deonstrated tht th ensmble delta confirming m ae emodel outperforms individual models and proved to be more suitable for lithology identification.
Abstract Faulted reservoirs are key targets for geothermal resources as well as lithium extraction, but their exploration at great depth faces great technological challenges. Recent advancements in drone and magnetic sensors have enabled the integration of measuring systems on light drones, allowing a great increase in the resolution of measurements. This study presents the successful execution of a very-high-resolution aeromagnetic survey (VHRAM) at 50 m above ground level using drones. The results were compared with regional interpretations of the Upper Rhine Graben from a helicopter-borne survey and seismic data. The drone-based survey provided a significant increase in resolution, enhancing the interpretation of local geologic features, particularly the basement rock structures and the prerift and synrift sediments in the fractured areas along the basin borders. A key finding was the ability to detect structural features in the basement at depths of several kilometers that were not observable in the regional survey and interpretation from seismic data. This very-high-resolution data, when combined with a reinterpretation of historical seismic lines, allowed refinement of the geologic model, highlighting a structural node in the area. Although the unmanned aerial vehicle (UAV) survey covered a relatively small area, its impact was significant, offering valuable insights at the scale of typical exploration permits. These results validated the feasibility and effectiveness of very-high-resolution aeromagnetic surveys (VHRAM) using lightweight drones at low altitudes for fault network characterization in geothermal resource exploration.
Abstract The long-term growth of a continental-shelf delta records combined sea-level fluctuations and sediment supply, by mixed hydrodynamic processes; consequently, its morphology typically features complex asymmetry. Using extensive 3D seismic data, cores, well logs, and provenance data, we present a systematic anatomy of the Early–Middle Miocene Paleo-Pearl River Delta (PPRD). Results show that fifteen depositional sequences can be identified, spanning T60 (~23 Ma) to T35 (~13.8 Ma). Each sequence exhibits a typical four-fold systems tract structure, from each showing typical deltaic planform morphologies and sedimentary facies. Mineralogical data reveal a distinct two-source pattern: the western branch mainly supplied the Enping Sag, whereas the eastern branch fed the Xijiang and Huizhou sags. This pattern is corroborated by 2D and 3D seismic data. Seismic geomorphology further shows an asymmetric morphology, with river channels and mouth bars deflected southwestward and beach and shelf sand ridges widening in the same direction. This asymmetry records a dominant southwestward paleoflow. Considering modern hydrodynamics and regional paleoclimatic evidence, we propose that monsoon-driven longshore currents and the Guangdong Coastal Current (GDCC) likely generated this flow. Reconstructing this paleocurrent system improves our understanding of sediment dynamics and reservoir prediction in the PPRD and comparable marine deltas.
Abstract Critical minerals (CM) are essential for emerging technologies; however, their exploration remains challenging because they occur in complex geological settings and often exhibit weak seismic expression. Volcanic facies (VF) are one of their sources of occurrence, where minerals such as lead, cobalt, copper, and nickel often accumulate and scatter due to hydrothermal processes. This study presents a facies-specific machine learning (ML) workflow that integrates a multi-attribute approach with supervised and unsupervised ML, aiming to resolve VF in a complex geologic setting in the Otway Basin, South Australia. Twenty-two seismic attributes were calculated from which facies-specific attribute combinations were selected. Attributes were then used as input into Self-Organizing Maps (SOM) and Independent Component Analysis (ICA) to find sills, lava flows, and dykes, while Probabilistic Neural Networks (PNN) was applied for supervised classification of pyroclastics, mounds, and laccoliths that were not identified using unsupervised methods. Results indicate that SOM better discriminates between sills and lava and effectively identifies dyke geometry. ICA improved the delineation of dyke boundaries and enhanced the contrast between the intrusions and the background medium, revealing features that were less distinct in SOM results. PNN successfully resolved complex features such as laccoliths. Our methodology demonstrates that specific attribute combinations processed through SOM, ICA, and PNN can effectively discriminate between closely related volcanic features that are structurally favorable for CM accumulation.
