Integrating OBN seismic data and machine learning for enhanced fluid discrimination in pre-salt carbonate reservoirs

Accurately identifying reservoir fluids in ultra-deep pre-salt reservoirs remains challenging due to complex lithology, strong heterogeneity, and limited well data. This study develops an integrated workflow combining ocean bottom node (OBN) seismic with pre-stack inversion, AVO analysis, and an LSTM rock physics model to improve fluid prediction. The methodology first extracts P-wave and S-wave impedances through pre-stack inversion, converting these into quantitative fluid indicators for distinguishing hydrocarbons, water zones, and non-reservoir facies. AVO cross-plotting enhances discrimination by exploiting distinct intercept-gradient patterns between oil and brine reservoirs. A key innovation replaces conventional shear-wave modeling with an LSTM network that autonomously learns nonlinear relationships among mineralogy, fluid distribution, and petrophysical parameters without empirical assumptions. Tests using synthetic and field datasets demonstrate superior performance in resolving ambiguous elastic responses characteristic of complex pre-salt reservoirs. The framework significantly reduces fluid identification uncertainties in frontier exploration areas where traditional methods fail, providing critical insights for hydrocarbon exploration while bridging the gap between rock physics theory and practical reservoir evaluation. This data approach offers a robust solution for reservoir characterization in geologically challenging environments, advancing exploration strategies through enhanced diagnostic accuracy.

1 Introduction

The identification of fluid properties in ultra-deep pre-salt continental limestone reservoirs represents a significant technical challenge for several reasons (Zhou et al., 2021; Valdez et al., 2025). These carbonate formations exhibit intricate pore network architectures, pronounced spatial variability in rock properties, and frequently suffer from insufficient well coverage to adequately constrain subsurface models (Eiken et al., 2003). Conventional marine seismic acquisition systems employing towed streamer configurations often prove inadequate for robust fluid prediction in such complex geological settings, primarily due to their restricted azimuthal sampling range and vulnerability to coherent noise interference (Zerafa et al., 2025; Zeng et al., 2023).

Recent technological developments in seafloor seismic acquisition have introduced ocean bottom node (OBN) systems that offer distinct advantages for reservoir characterization (Smit et al., 2008). These advanced acquisition methods provide comprehensive azimuthal coverage, superior signal fidelity, and the capability to record full elastic wavefield information (Zhang W. et al., 2021; Yang et al., 2023). Such characteristics make OBN surveys particularly valuable for pre-salt exploration programs, where precise fluid differentiation constitutes a crucial factor in mitigating operational risks associated with exploratory drilling (Da Silva et al., 2024; França et al., 2021).

Progress in seismic inversion methodologies and rock physics modeling techniques has led to more accurate estimation of reservoir parameters (Shakir et al., 2024). Contemporary approaches frequently employ pre-stack simultaneous inversion to derive key elastic properties - including P-wave impedance, S-wave impedance, and bulk density for improved fluid discrimination (Du et al., 2019). Complementary amplitude variation with offset (AVO) analysis further enhances fluid identification by capitalizing on characteristic seismic response patterns, where hydrocarbon-bearing intervals typically display diagnostic trends in intercept-gradient crossplots that differ markedly from water-saturated zones (Veeken and Rauch-Davies, 2006). However, established rock physics frameworks such as the Xu-Payne model and differential effective medium theory often encounter difficulties in reliably predicting S-wave velocity in continental carbonates, owing to their complex pore geometry characteristics and post-depositional diagenetic modifications (Sun et al., 2012). These limitations introduce substantial uncertainties in the derivation of fluid-sensitive seismic attributes, particularly in pre-salt reservoir contexts where conventional empirical correlations may prove unreliable (Ariza et al., 2021; Li S. et al., 2019).

Emerging machine learning approaches, especially those utilizing deep neural networks, have demonstrated considerable potential in overcoming these challenges (Bertolini et al., 2021; Hernández Medina et al., 2022). Recurrent neural network architectures like long short-term memory (LSTM) networks have shown notable success in predicting S-wave velocities from petrophysical log data by effectively modeling complex nonlinear relationships among mineral composition, pore space characteristics, and elastic moduli (You et al., 2021). The integration of such methodologies with physics-based inversion techniques may provide a pathway to reconcile seismic response characteristics with subsurface fluid properties (Xu and Heagy, 2025; Wu et al., 2023). Despite these advances, current research efforts have not fully capitalized on the multi-azimuth capabilities of OBN data when combined with artificial intelligence-enhanced rock physics modeling for pre-salt fluid characterization (Zhang Y. et al., 2021).

This investigation presents an innovative methodological framework that synthesizes OBN seismic data analysis, pre-stack inversion procedures, AVO attribute evaluation, and LSTM-enabled rock physics modeling to enhance fluid discrimination in ultra-deep pre-salt continental limestone reservoirs. The proposed technique harnesses the comprehensive azimuthal coverage provided by OBN surveys to improve the stability of elastic parameter inversion, while AVO crossplot analysis offers supplementary constraints for fluid type classification. A critical methodological advancement involves replacing conventional S-wave velocity prediction models with an LSTM network trained on comprehensive well log datasets, thereby reducing dependence on empirical approximations and enhancing predictive performance for previously un-encountered reservoir conditions. The synergistic combination of these components facilitates more confident fluid identification in geologically complex scenarios where traditional methodologies produce ambiguous interpretations.

The developed approach makes substantive contributions to the field through three principal innovations. First, it systematically integrates the distinctive advantages of OBN acquisition, including multi-azimuth angle gathers and converted wave information, into the derivation of fluid-sensitive seismic attributes, thereby addressing inherent limitations associated with conventional streamer data. Second, it implements an LSTM-based rock physics modeling framework capable of adaptively learning complex relationships between reservoir properties and seismic response characteristics, circumventing biases introduced by empirical modeling approaches. Third, the methodology provides a practical operational solution for exploration scenarios characterized by sparse well control, where traditional rock physics calibration procedures often prove inadequate.

2 Methods

The characterization of subsurface reservoir fluids utilizing seismic data has undergone substantial transformation, propelled by innovations in acquisition methodologies and computational algorithms. Traditional approaches predominantly depended upon post-stack attribute analysis and heuristic correlations linking seismic amplitude variations with fluid composition (Chopra and Marfurt, 2007). However, these techniques exhibited pronounced limitations in structurally intricate geological formations, primarily due to their failure to incorporate amplitude variations across offset and azimuthal dependencies. A paradigm shift occurred with the advent of pre-stack inversion methodologies, facilitating the direct derivation of elastic parameters from seismic gathers (Jianguo and Ntibahanana, 2024).

2.1 Advancements in OBN seismic applications

As hydrocarbon exploration ventures into increasingly deeper offshore basins, the necessity for enhanced precision in fluid identification within deep and ultra-deep reservoirs has intensified. Traditional marine streamer seismic surveys, characterized by narrow-azimuth data acquisition, have proven inadequate for contemporary exploration demands. Consequently, alternative methodologies such as Ocean Bottom Cable (OBC) and Ocean Bottom Node (OBN) systems have been progressively integrated into marine seismic operations.

Conventional 3D marine acquisition relies on single-vessel, dual-source multi-streamer configurations—a design inherently limited to narrow-azimuth pseudo-3D observation (Xue et al., 2024). Expanding azimuthal coverage in streamer surveys necessitates complex multi-vessel coordination and repeated survey line transits, leading to operational inefficiencies (Zhang and Luo, 2021). In contrast, OBN technology circumvents these constraints through seabed-deployed autonomous nodes, which encapsulate seismic receivers. Nodes are precisely positioned at predetermined locations, while source vessels can operate across unrestricted azimuths within the node deployment area, thereby enabling highly efficient wide-azimuth, high-fold data acquisition (Figure 1) (Li B. et al., 2019; Wu et al., 2021).

Figure 1

Figure 1. Ocean bottom node (OBN) seismic survey system.

OBN acquisition entails significantly greater costs compared to streamer technology. However, the exceptionally high-quality data it provides makes it a worthwhile investment for addressing specific geological problems and reservoir monitoring in challenging environments. Recent progress in seismic acquisition instrumentation has markedly enhanced OBN operational efficiency. Concurrently, innovations in electromechanical engineering and evolving commercial paradigms have substantially reduced associated costs. These dual advances have catalyzed sustained expansion in OBN applications (Zhang et al., 2019; Ronholt et al., 2008; Mienert et al., 2005; Bovet et al., 2010; Reasnor et al., 2010; Stone et al., 2018). Following initial field trials in North American hydrocarbon provinces aimed at obtaining long-offset, full-azimuth data, OBN technology has been globally implemented, demonstrating significant advantages for offshore exploration. Although China’s adoption of OBN exploration commenced later relative to international counterparts, current operations demonstrate notable progress.

OBN seismic datasets exhibit superior characteristics compared to conventional streamer data, including expansive azimuthal coverage, enhanced signal-to-noise ratios, reduced interference, and richer angular information. These attributes facilitate more robust reservoir property predictions through inversion analyses, particularly for ultra-deep offshore reservoirs requiring detailed fluid characterization.

The multi-component nature of OBN systems permits simultaneous recording of compressional (P-wave) and converted-wave (PS) energy—a capability empirically demonstrated to enhance reservoir characterization in complex geological settings (Su et al., 1998; Tellier and Herrmann, 2023). Emerging research highlights OBN’s efficacy in challenging environments such as subsalt imaging, where conventional methods encounter difficulties due to intricate wave propagation pathways (Blanch et al., 2020). However, applications specifically targeting pre-salt carbonate reservoirs with complex pore geometries remain an area requiring further investigation.

2.2 Pre-stack elastic simultaneous inversion method based on OBN seismic data

Seismic inversion constitutes a fundamental methodology for subsurface reservoir characterization. Contemporary inversion approaches bifurcate into post-stack and pre-stack methodologies. Post-stack inversion techniques have evolved substantially over three decades, progressing from rudimentary linear inversions to sophisticated nonlinear and depth-domain implementations (Zhang et al., 2010).

Pre-stack simultaneous inversion leverages partial angle stack volumes and approximations of the Zoeppritz equations to derive petrophysically relevant parameters including P-wave impedance (Ip), S-wave impedance (Is), density (ρ), and Vp/Vs ratios. This research employs elastic parameter simultaneous inversion, capitalizing on the rich angle-dependent information inherent in pre-stack data. The incorporation of Zoeppritz physics substantially enhances inversion fidelity, generating comprehensive elastic property estimates requisite for fluid discrimination.

Comparative analyses were conducted between conventional streamer data, OBN datasets, and multi-azimuth OBN acquisitions. As illustrated in Figure 2, the principal advantage of OBN data manifests in the availability of extended-offset angle gathers included in the inversion workflow.

Figure 2

Figure 2. Flowchart of pre-stack elastic parameter inversion.

2.3 AVO attribute analysis method based on OBN seismic data

The AVO technique focuses on analyzing amplitude variations with offset within angle gathers to delineate lithological characteristics and fluid distributions. This approach generates multiple diagnostic profiles, including intercept (P), gradient (G), pseudo-Poisson’s ratio, fluid indicators, density variations, and P-wave/S-wave reflectivity coefficients. The primary objective of this analysis is to establish meaningful correlations between seismic amplitude responses and subsurface reservoir properties through seismic petrophysical relationships.

A robust AVO framework necessitates comprehensive integration of geological knowledge with geophysical observations. The process involves constructing reliable AVO diagnostic criteria by synthesizing geological models, well log data, drilling records, and seismic responses. This integrated approach enables optimal utilization of AVO information for reservoir characterization. The current investigation employs the Aki-Richards three-term approximation to derive fundamental AVO parameters, particularly the intercept (P) and gradient (G) attributes, which are subsequently combined in various configurations to analyze fluid content in reservoir units.

The intercept attribute exhibits superior zero-offset approximation characteristics compared to conventional CRP data, providing more reliable amplitude information for reservoir property estimation. In contrast, the gradient attribute reflects the composite elastic response of subsurface formations by capturing variations in P-wave velocity, S-wave velocity, and bulk density. Interpretation of gradient polarity requires careful consideration of event polarity: for peak events (positive reflectivity), positive gradients indicate amplitude attenuation with increasing offset, while negative gradients suggest the opposite trend; for trough events (negative reflectivity), this relationship reverses as detailed in Table 1. Therefore, comprehensive AVO mandates joint analysis of both intercept and gradient attributes.

Table 1

Table 1. The physical significance of P × G.

When the P-wave to S-wave velocity ratio (Vp/Vs) approaches 2, the P-G crossplot effectively characterizes shear impedance properties, functioning as a pseudo-shear wave profile. Under these specific velocity ratio conditions, the combined P + G parameter can be interpreted as a proxy for Poisson’s ratio. Furthermore, the product of intercept and gradient (P × G) serves as a robust hydrocarbon indicator, with anomalously high values typically corresponding to hydrocarbon-bearing zones, thus providing a reliable fluid detection metric in AVO analysis.

AVO forward modeling utilizes model-based simulation to replicate AVO phenomena. This process analyzes the AVO responses of oil, gas, water, and special lithologies by integrating local reservoir characteristics. It establishes identification criteria to enable the direct detection of lithology and hydrocarbons in seismic data. If there is a certain understanding of the relationship between reservoir porosity and seismic lithological parameters, it can also qualitatively identify pore development zones. Based on the difference in wave impedance between the reservoir and surrounding rocks, AVO types can be classified into four categories (Figure 3) (Castagna and Swan, 1997).

Figure 3

Figure 3. The classification of AVO types.

Class I: High-impedance reservoir. Exhibits strong zero-offset amplitude (positive polarity). The amplitude decreases as the incidence angle increases, eventually leading to a polarity reversal. Often manifests as a “dim spot” on stacked sections.

Class II: Near-equal impedance. Characterized by weak zero-offset amplitude (approaching zero). The absolute amplitude value increases significantly as the incidence angle increases (negative gradient), potentially forming a “bright spot”.

Class III: Low-impedance reservoir. Exhibits strong zero-offset amplitude (negative polarity). The absolute amplitude value continues to increase as the incidence angle increases (negative gradient), very easily forming a “bright spot”.

Class IV: Low-impedance reservoir. Exhibits strong zero-offset amplitude (negative polarity). However, the absolute amplitude value decreases as the incidence angle increases (positive gradient).

2.4 Machine learning in rock physics

Recent advances in computational methods have facilitated the emergence of machine learning as a transformative approach for resolving nonlinear correlations between reservoir characteristics and seismic response patterns. Initial explorations employed artificial neural networks to estimate S-wave velocities from standard well logging measurements (Maier and Dandy, 2000). Contemporary developments have seen convolutional neural networks (CNNs) demonstrate enhanced performance in seismic inversion tasks compared to conventional optimization techniques (Li et al., 2020). Particularly noteworthy are long short-term memory (LSTM) networks, which have exhibited substantial potential for petrophysical parameter estimation from sequential well log data, owing to their capacity to model temporal dependencies (Li and Gao, 2023). However, current applications predominantly concentrate on clastic reservoirs, while carbonate formations–where mineral heterogeneity and diagenetic alterations significantly influence elastic properties–remain comparatively understudied.

The significance of S-wave velocity measurements extends beyond basic reservoir characterization, serving as a critical parameter for comprehensive fluid prediction analyses. When integrated with P-wave velocities, these measurements provide essential insights into rock mechanical properties, reservoir architecture, and fluid dynamics (Buland and Omre, 2003). Practical constraints in hydrocarbon exploration, particularly cost considerations associated with extensive well logging campaigns, frequently result in sparse S-wave data acquisition. This data scarcity presents substantial obstacles for accurate pre-stack seismic interpretation and reservoir prediction, as conventional inversion methodologies typically require complete elastic parameter datasets. Consequently, robust rock physics modeling frameworks become indispensable for reliable reservoir evaluation under such data limitations.

2.4.1 Traditional rock physics modeling methods

Established techniques for S-wave velocity prediction can be broadly classified into two methodological paradigms: empirical correlation methods and physical modeling approaches. While empirical formulations offer computational simplicity and rapid implementation, their predictive validity remains geographically constrained and often yields suboptimal accuracy. Physical modeling methods, conversely, derive S-wave velocities through fundamental physical principles involving mineralogical composition, porosity distribution, and fluid phase properties. Although theoretically more rigorous, these approaches encounter practical deployment challenges stemming from parameter complexity, computational intensity, and the requisite interdisciplinary expertise for proper implementation.

The Xu-White theoretical framework, grounded in dual-phase medium and effective medium theories, conceptualizes rock systems through four constituent elements: mineral matrix, dry rock framework, pore constituents, and fluid-saturated rock. This comprehensive model adopts an inclusive definition of porosity that encompasses conventional pore spaces, structures, and fracture networks (Ying and Liu, 2016; Cai et al., 2013). Figure 4 systematically presents the computational workflow inherent to the Xu-White model. While theoretically robust, the methodology’s practical implementation proves cumbersome due to its parametric complexity and computationally intensive sequential calculation procedures.

Figure 4

Figure 4. Flowchart of the Xu-White model rock physics modeling methodology.

2.4.2 Rock physics modeling method based on LSTM network

The proliferation of artificial intelligence applications has catalyzed significant developments in hydrocarbon exploration data processing and predictive modeling. Figure 5 delineates the structural workflow of the LSTM-based rock physics modeling approach. The methodology involves feeding mineralogical data and comprehensive well log measurements into a purpose-designed neural network architecture. Through iterative machine learning processes, the system autonomously establishes correlations between mineral constituents and logging responses. During operational deployment, input parameters including mineral fractions, well log, and fluid content enable the trained network to efficiently predict S-wave velocities (Figure 6). These input data are actual measured logging data. The LSTM model was configured with the hyper-parameters (learning rate = 0.001, epochs = 200, validation split = 0.2, patience = 5) to train the network. To achieve this trained network, the most computationally intensive steps were performed on a standard computing node equipped with an NVIDIA RTX 3080 GPU, 64 GB of RAM, and 12-core CPU. A single training run typically required 2–3 h (highly dependent on parameters and the number of iterations), while generating S-wave velocity predictions took approximately 3–4 min.

Figure 5

Figure 5. Flowchart of the rock physics modeling methodology based on LSTM network.

Figure 6

Figure 6. Schematic diagram of rock physics modeling based on LSTM network.

In this study, a deep learning model based on LSTM network was employed for rock physics modeling, specifically to predict the velocities. The network comprises an LSTM layer and a dense layer. LSTM is a special type of Recurrent Neural Network (RNN) capable of processing and predicting sequential data. By introducing a gating mechanism (forget gate, input gate, output gate), it addresses the long-term dependency issue inherent in traditional RNNs. This enables a better understanding of information at different positions within sequential data. The LSTM layer contains 256 hidden units and utilizes the default tanh activation function. This activation function aids in handling complex nonlinear relationships. After the LSTM layer, the data flows through a dense layer, which maps the features extracted by the LSTM to the final output. In the dense layer, we used the ReLU activation function. ReLU effectively helps the network learn complex patterns and offers advantages such as high computational efficiency and fast convergence (Figure 7). The loss function is a metric that measures the discrepancy between the model’s predictions and the actual values. For this regression task, we selected Mean Squared Error (MSE) as the loss function. MSE calculates the squared difference between the predicted and actual values, with the mathematical formula as follows:

$MSE=\frac{1}{N}{\displaystyle \sum _{i=1}^N}{\left(y_i-{\widehat{y}}_i\right)}^2$

$N$ is the number of samples, $y_i$ is the actual values, ${\widehat{y}}_i$ is the predicted values. By minimizing the MSE, the model adjusts its parameters to minimize the error between predicted values and actual values. This helps improve the predictive accuracy of the model. Comparing the conventional rock physics modeling applying the Xu-White model with the LSTM-based rock physics modeling, we found that the latter achieved a prediction error of 2.77%, significantly reduced compared to the former 12.7% (Figure 8). This innovative methodology introduces several substantive advances over conventional techniques. Primarily, it capitalizes on the multi-azimuth illumination properties of OBN data to enhance elastic parameter inversion stability in complex pre-salt geological settings. Secondly, the approach supplants traditional empirical rock physics models with an LSTM framework capable of autonomously learning intricate relationships among mineral composition, porosity, and elastic properties from available well data. Thirdly, the integrated workflow systematically combines multiple fluid-sensitive attributes, thereby mitigating interpretation uncertainties inherent to single-attribute analyses. This synergistic combination of advanced acquisition technology, data modeling, and multi-parametric analysis represents a paradigm shift in pre-salt reservoir characterization methodologies.

Figure 7

Figure 7. The mechanism of LSTM model.

Figure 8

Figure 8. The comparison of predicted Vs. and actual Vs. based on LSTM rock physics modeling and conventional rock physics modeling.

3 Results and discussion

The present investigation centers on the S Oilfield located within Brazil’s Santos Basin, a representative ultra-deep pre-salt carbonate reservoir exhibiting intricate lithological variations and heterogeneous fluid distribution patterns. This comprehensive analysis incorporates OBN (Ocean Bottom Node) seismic acquisition data obtained through a sophisticated 4-component (4C) recording system, delivering full-azimuth illumination (0°–360°) with maximum offsets reaching 13 km. For benchmarking purposes, conventional towed-streamer seismic data from the same geological setting, characterized by more limited offsets (up to 8 km), were included in the comparative analysis. The study leverages calibration data from 15 strategically placed exploration and production wells (notably Y1, Y7, and Y13), targeting a reservoir interval spanning 5000–6000 m depth beneath substantial salt layers that induce pronounced seismic signal attenuation. The reservoir units comprise microbial carbonate facies with porosity values distributed between 4% and 30% and permeability exhibiting significant heterogeneity (0.1–2000 mD).

3.1 Fluid prediction methodology in sub-salt carbonate reservoirs

3.1.1 Feasibility analysis of fluid prediction using OBN seismic

To address seismic imaging challenges in the S Oilfield, an extensive OBN seismic survey was implemented. The study area contains two distinct ultra-deep continental limestone reservoirs beneath thick salt layer. Our investigations reveal that the overlying salt formations severely attenuate seismic energy, leading to compromised subsalt resolution. Furthermore, the presence of igneous intrusions within the reservoir interval introduces additional complexities for fluid characterization. These technical constraints necessitate the application of advanced acquisition techniques such as OBN to facilitate efficient field development.

Spectral analysis of pre-stack versus post-stack seismic data demonstrates noticeable frequency attenuation within the 5000–6000 m depth interval, attributable to wavefront divergence and stratigraphic absorption effects. Quantitative measurements indicate post-stack data exhibits a 9 Hz dominant frequency with 3–18 Hz bandwidth, while pre-stack data shows superior resolution with 11 Hz dominant frequency and 3–23 Hz bandwidth (Figure 9). This enhanced frequency content in pre-stack data establishes pre-stack inversion as the preferred methodology for fluid discrimination in this geological setting.

Figure 9

Figure 9. (a) OBN seismic data acquisition coverage map of the S oilfield development area; (b) Comparison of spectral information between post-stack and pre-stack seismic data in the reservoir.

3.1.2 Fluid sensitivity factor development

Reservoir fluid characterization requires integrated interpretation of diverse datasets including well logs, seismic attributes, and geological constraints. Our technical approach involved systematic evaluation of multiple elastic parameter combinations (P-wave impedance, S-wave impedance, Vp/Vs, and density) to derive optimal fluid indicators (Figure 10). Through rigorous parameter analysis, we established the product of P-wave and S-wave impedances ($I_p{I}_s$) as the most effective fluid-sensitive factor F for this reservoir system, mathematically expressed as:

$F=I_p{I}_s$

Figure 10

Figure 10. Cross-plot analysis of rock physics parameters. (a) Vp versus Vs.; (b) Impedance P versus Impedance S; (c) Density versus Vp/Vs; (d) Porosity versus Vp; (e) Fluid factor λp versus Vp/Vs; (f) Fluid factor F versus Vp/Vs.

Cross-plot analysis of petrophysical parameters, constrained by well data, enabled clear discrimination of fluid zones (Figure 10f). The derived classification criteria are: oil zones (0.5 × 10⁸–1 × 10⁸), water zones (1 × 10⁸–1.4 × 10⁸), and non-reservoir zones (1.4 × 10⁸–2.0 × 10⁸). These quantitative thresholds provide reliable benchmarks for reservoir evaluation.

3.1.3 AVO response characteristics

Preprocessed OBN pre-stack gathers were subjected to comprehensive AVO analysis to establish fluid discrimination criteria. Forward modeling results from 15 wells demonstrate that reservoir zones consistently exhibit lower impedance contrasts compared to non-reservoir. Our findings reveal systematic AVO behavior differences between fluid phases: oil reservoirs display Class I responses (P > 0, G < 0) with characteristic amplitude decay versus offset (e.g., Well Y13), while water zones consistently show Class IV signatures (P < 0, G < 0) as exemplified by Well Y8 (Figure 11).

Figure 11

Figure 11. Schematic diagram of AVO forward modeling for oil and water zones in the reservoir.

The AVO modeling results from principal study areas demonstrate distinct seismic response patterns: hydrocarbon reservoirs predominantly display Class I AVO characteristics (G < 0, P > 0), whereas water-saturated formations show Class IV responses (G > 0, P > 0). These intercept and gradient profiles were derived through rigorous application of the Aki-Richards three-term approximation.

Analysis of the combined P + G attribute reveals significant limitations, as both oil (P_oil>0, G_oil<0) and water (P_water<0, G_water>0) formations yield mathematically similar results, rendering this parameter ineffective for fluid differentiation. In contrast, the P-G crossplot demonstrates better discrimination capacity, with oil zones consistently plotting above zero (P_oil-G_oil>0) while water zones show negative trends (P_water-G_water<0), as visually confirmed in Figure 12. Notably, the anomalous negative values (yellow) observed at the upper section of Well Y1’s target formation correspond to impermeable diabase intrusions rather than water zones. Productive intervals in Wells Y1 (middle section) and Y7 (lower section) exhibit positive P-G attributes (blue), corroborating well-log interpretations. However, the subtle positive response (blue) from Well Y13’s productive upper section suggests this parameter has variable predictive capability across different wells.

Figure 12

Figure 12. Well-Tie Profiles of OBN Pre-stack Seismic AVO. (a) Pseudo-Poisson’s Ratio 4/3(P + G); (b) Pseudo-S-wave impedance 1/2(P-G); (c) Hydrocarbon detection P × G.

For hydrocarbon detection, the P × G attribute reveals distinct yellow negative values in the Y1 and Y7 oil reservoirs, while the Y13 reservoir displays only intermittent negatives. This confirms that P × G is partially effective in identifying oil, but unable to separate it from water.

This method uses pre-stack seismic data, but lacks other constraints. The attribute values range between [-1, 1], with a limited scope and no physical significance in relation to rock elastic parameters. Therefore, while it is somewhat rational for oil-water identification, it can only serve as the basis for a qualitative analysis of oil-water distribution characteristics. Consequently, while providing reasonable qualitative insights into fluid distribution patterns, this approach cannot support quantitative fluid characterization and should be integrated with additional geological and geophysical constraints for comprehensive reservoir evaluation.

3.1.4 Pre-stack simultaneous inversion and fluid identification

Elastic parameter simultaneous inversion was performed on near-, mid-, and far-angle stacked OBN seismic data. Through key steps such as S-wave fitting, gather optimization, wavelet extraction, and well-seismic calibration, inversion data including P-wave impedance, S-wave impedance, and density were obtained under the constraint of a refined low-frequency model. Based on the formula of fluid factor F, inversion results of fluid factor F for azimuth angles (0°, 60°, 120°) were derived (Figure 13). In the profile, red low values indicate oil, green and cyan indicate water, and purple high values indicate non-reservoirs. In the 0° azimuth data profile, red low values appear in the BVE formation of wells Y1, Y7, and Y13, representing oil, while the bottom of the ITP formation shows cyan water, consistent with well-log interpretations. In the 60° azimuth profile, the oil match at well Y1 is slightly weaker compared to the 0° azimuth profile. In the 120° azimuth profile, the non-reservoir in the upper BVE formation of wells Y1and Y7 are more pronounced, and the inversion profiles align better with well-log curves at all well points.

Figure 13

Figure 13. Inversion results of Pre-stack fluid factor F. (a) 0° Azimuth inversion profile; (b) 60° Azimuth inversion profile; (c) 120° Azimuth inversion profile.

For plane analysis, RMS amplitudes were extracted along layers from the fluid factor F results (Figure 14). By comparing with well-drilled reservoir conditions, it was concluded that the inversion of OBN seismic data matched 13 out of 15 wells, with a compliance rate of 87%, while the towed-streamer seismic inversion matched 10 out of 15 wells, with a compliance rate of 67%. The fluid prediction compliance rate of OBN seismic data improved by 20%.

Figure 14

Figure 14. Comparison of oil-water prediction distribution. (a) Oil-water prediction distribution in OBN seismic data; (b) Oil-water prediction distribution in towed-streamer seismic data.

The fluid factor F = 0.5 × 10⁸–1 × 10⁸ was defined as the oil indicator criterion. Using this fluid factor F, 3D oil thickness maps were constructed for the BVE and ITP formations, as shown in Figure 15. Wells Y1 and Y6, located in the paleo-uplift area, exhibit relatively thick oil layers exceeding 200 m, whereas wells Y4 and Y7 show thinner layers. Well Y13 has the thinnest oil layer and the poorest physical properties, which is consistent with well-log interpretation data.

Figure 15

Figure 15. 3D map of oil-layer thickness in the target zone.

Based on pre-stack simultaneous inversion profiles of OBN seismic data from different azimuth angles, the fluid factor F inversion profiles show high consistency with well-log results from each well, while P-wave impedance, S-wave impedance, and density exhibit slightly lower sensitivity to reservoir fluid identification in the study area. The minor uncertainties that persist in local oil-water discrimination are primarily linked to the influence of igneous intrusions. Nevertheless, the technique manifests significantly higher reliability than AVO attribute analysis for reservoir characterization. The method’s effectiveness is particularly evident in its consistent alignment with well-based interpretations across multiple azimuthal datasets.

4 Future work

Despite its advantages, the proposed workflow faces several practical limitations. The requirement for high-quality OBN data introduces acquisition cost considerations, particularly in frontier exploration areas where well control is sparse. Quality control and uncertainty assessment were conducted on the input well log data. However, the inherent uncertainties in these datasets remain potential of residual error in the results, particularly affecting the LSTM training and prediction. Processing multi-azimuth angle gathers demands significant computational resources, especially when incorporating full-waveform inversion techniques for improved resolution (Zhang et al., 2020). Furthermore, the predictive reliability of the LSTM-based rock physics model remains contingent upon the comprehensiveness of training datasets, potentially leading to estimation inaccuracies when applied to reservoir formations exhibiting substantially distinct diagenetic evolution pathways.

The methodology’s core principles extend beyond pre-salt carbonates to other complex reservoirs where conventional fluid indicators prove inadequate. Tight gas sandstones with low porosity-permeability characteristics exhibit similar challenges in S-wave velocity prediction, suggesting potential for LSTM adaptation in such settings (Guo et al., 2023). Integrating electromagnetic (EM) data could further enhance discrimination in high-salinity formation water environments, where resistivity contrasts complement seismic attributes (Katterbauer et al., 2014). Future implementations may also incorporate distributed acoustic sensing (DAS) measurements from borehole fiber optics, providing direct constraints on wave propagation anisotropy near wellbores (Chavarria, 2018).

5 Conclusion

The proposed methodology demonstrates significant advancements in fluid identification for ultra-deep pre-salt carbonate reservoirs by leveraging the unique advantages of OBN seismic data combined with machine learning-enhanced rock physics modeling. The integration of multi-azimuth pre-stack inversion, fluid-sensitive factor F, and LSTM-based S-wave velocity prediction addresses critical limitations of conventional approaches in complex carbonate settings. Key findings reveal that the fluid factor F achieves superior fluid discrimination compared to traditional attributes, while the LSTM network reduces S-wave velocity prediction errors relative to empirical models.

The 120° azimuth profile emerges as the most reliable for fluid prediction due to its perpendicular orientation to the dominant fracture direction in the geological formation. While other azimuths provide valuable complementary information when integrated through the multi-azimuth framework. The method’s robustness is further validated by its consistent performance across varying porosity ranges and complex pore geometries characteristic of microbial carbonate reservoirs.

These results have immediate practical implications for exploration risk reduction in pre-salt provinces, where accurate fluid identification prior to drilling carries substantial economic consequences. The approach provides a template for adapting advanced acquisition technologies with rock physics in other geologically complex settings, from tight gas sandstones to fractured basements. Future refinements should focus on optimizing computational efficiency for routine application.

The successful field application in the Santos Basin establishes a new benchmark for reservoir characterization in salt-proximity environments, demonstrating how modern geophysical techniques can overcome longstanding limitations in seismic fluid discrimination. As the energy industry increasingly targets challenging reservoirs, such integrated approaches will become indispensable tools for reducing uncertainty and improving exploration success rates.

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Author contributions

SW: Writing – original draft, Data curation. CW: Visualization, Writing – review and editing. GF: Supervision, Writing – review and editing. HW: Writing – review and editing. GZ: Writing – review and editing. LD: Writing – review and editing. YZ: Writing – review and editing. XP: Writing – review and editing. QL: Writing – review and editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication.

Conflict of interest

Authors SW, CW, GF, HW, GZ, LD, YZ, and XP were employed by PetroChina Hangzhou Research Institute of Geology.

Author QL was employed by CNPC Bohai Drilling Engineering Company No. 1 Mud Logging Company.

The author(s) declared that this work was supported by the CNPC Research on Key Technologies for Overseas Deepwater Oil and Gas Field Exploration and Development Projects (2023-SC-01-03). The funder had the following involvement in the study: decision to publish.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart.2026.1767668/full#supplementary-material

References

Ariza, F. D. J., Dias, R. M., and Lupinacci, W. M. (2021). Seismic pattern classification integrated with permeability-porosity evaluation for reservoir characterization of presalt carbonates in the buzios field, Brazil. J. Petroleum Sci. Eng. 201, 108441. doi:10.1016/j.petrol.2021.108441

CrossRef Full Text | Google Scholar

Bertolini, M., Mezzogori, D., Neroni, M., and Zammori, F. (2021). Machine learning for industrial applications: a comprehensive literature review. Expert Syst. Appl. 175, 114820. doi:10.1016/j.eswa.2021.114820

CrossRef Full Text | Google Scholar

Blanch, J., Jarvis, J., Hurren, C., Kostin, A., Liu, Y., and Hu, L. (2020). Designing an exploration-scale OBN: acquisition design for subsalt imaging and velocity determination. Lead. Edge 39 (4), 248–253. doi:10.1190/tle39040248.1

CrossRef Full Text | Google Scholar

Bovet, L., Ceragioli, E., Tchikanha, S., Guilbot, J., and Toinet, S. (2010). “Ocean bottom nodes processing: reconciliation of streamer and OBN data sets for time lapse seismic monitoring. The Angolan deep offshore experience,” in SEG technical program expanded abstracts 2010 (Houston, TX: Society of Exploration Geophysicists), 3751–3755. Available online at: https://library.seg.org/doi/10.1190/1.3513630.

CrossRef Full Text | Google Scholar

Buland, A., and Omre, H. (2003). Bayesian linearized AVO inversion. Geophysics 68 (1), 185–198. doi:10.1190/1.1543206

CrossRef Full Text | Google Scholar

Cai, H. P., He, Z. H., Tang, X. R., He, G. M., and Zou, W. (2013). Influence analysis of carbonate pore structure and calculation of equivalent pore structure parameters. Geophys. Prospect. Petroleum 52 (6), 566–572. doi:10.3969/j.issn.1000-1441.2013.06.002

CrossRef Full Text | Google Scholar

Castagna, J., and Swan, H. W. (1997). Principles of AVO crossplotting. Houston, TX: Society of Exploration Geophysicists.

Google Scholar

Chavarria, J. (2018). “Reservoir monitoring through DAS measurements,” in 80th EAGE conference & exhibition 2018 workshop programme (Copenhagen, Denmark: European Association of Geoscientists & Engineers). Available online at: https://www.earthdoc.org/content/papers/10.3997/2214-4609.201801916.

Google Scholar

Chopra, S., and Marfurt, K. J. (2007). “Overview of seismic attributes,” in Seismic attributes for prospect identification and reservoir characterization (Houston, TX: Society of Exploration Geophysicists), 1–24. Available online at: https://library.seg.org/doi/10.1190/1.9781560801900.ch1.

Google Scholar

Da Silva, SLEF, Costa, F. T., Karsou, A., De Souza, A., Capuzzo, F., Moreira, R. M., et al. (2024). Refraction FWI of a circular shot OBN acquisition in the Brazilian presalt region. IEEE Trans. Geosci. Remote Sens. 62, 1–18. doi:10.1109/tgrs.2024.3426956

CrossRef Full Text | Google Scholar

Du, J., Liu, J., Zhang, G., Han, L., and Li, N. (2019). “Pre-stack seismic inversion using SeisInv-ResNet,” in SEG technical program expanded abstracts 2019 (San Antonio, Texas: Society of Exploration Geophysicists), 2338–2342. Available online at: https://library.seg.org/doi/10.1190/segam2019-3215750.1.

CrossRef Full Text | Google Scholar

Eiken, O., Haugen, G. U., Schonewille, M., and Duijndam, A. (2003). A proven method for acquiring highly repeatable towed streamer seismic data. Geophysics 68 (4), 1303–1309. doi:10.1190/1.1598123

CrossRef Full Text | Google Scholar

França, D., Coutinho, D. M., Barra, T. A., Xavier, R. S., and Azevedo, D. A. (2021). Molecular-level characterization of Brazilian pre-salt crude oils by advanced analytical techniques. Fuel 293, 120474. doi:10.1016/j.fuel.2021.120474

CrossRef Full Text | Google Scholar

Guo, Z. Q., Qin, X. Y., and Liu, C. (2023). Quantitative characterization of tight gas sandstone reservoirs using seismic data via an integrated rock-physics-based framework. Petroleum Sci. 20 (6), 3428–3440. doi:10.1016/j.petsci.2023.09.003

CrossRef Full Text | Google Scholar

Hernández Medina, R., Kutuzova, S., Nielsen, K. N., Johansen, J., Hansen, L. H., Nielsen, M., et al. (2022). Machine learning and deep learning applications in microbiome research. ISME Commun. 2 (1), 98. doi:10.1038/s43705-022-00182-9

PubMed Abstract | CrossRef Full Text | Google Scholar

Jianguo, S., and Ntibahanana, M. (2024). Developing deep learning methods for pre-stack seismic data inversion. J. Appl. Geophys. 222, 105336. doi:10.1016/j.jappgeo.2024.105336

CrossRef Full Text | Google Scholar

Katterbauer, K., Hoteit, I., and Sun, S. (2014). EMSE: synergizing EM and seismic data attributes for enhanced forecasts of reservoirs. J. Petroleum Sci. Eng. 122, 396–410. doi:10.1016/j.petrol.2014.07.039

CrossRef Full Text | Google Scholar

Li, J., and Gao, G. (2023). Digital construction of geophysical well logging curves using the LSTM deep-learning network. Front. Earth Sci. 10, 1041807. doi:10.3389/feart.2022.1041807

CrossRef Full Text | Google Scholar

Li, S., Wang, D., and Zhang, M. (2019). Influence of upscaling on identification of reservoir fluid properties using seismic-scale elastic constants. Sci. Rep. 9 (1), 13056. doi:10.1038/s41598-019-49559-2

PubMed Abstract | CrossRef Full Text | Google Scholar

Li, B., Feng, Q. K., Zhang, Y. B., and Huang, F. Q. (2019). Summary of development and key issues of offshore OBC-OBN technology. Geophys. Geochem. Explor. 43 (6), 1277–1284. doi:10.11720/wtyht.2019.0370

CrossRef Full Text | Google Scholar

Li, S., Liu, B., Ren, Y., Chen, Y., Yang, S., Wang, Y., et al. (2020). Deep-learning inversion of seismic data. IEEE Trans. Geosci. Remote Sens. 58 (3), 2135–2149. doi:10.1109/tgrs.2019.2953473

CrossRef Full Text | Google Scholar

Maier, H. R., and Dandy, G. C. (2000). Neural networks for the prediction and forecasting of water resources variables: a review of modelling issues and applications. Environ. Model. & Softw. 15 (1), 101–124. doi:10.1016/s1364-8152(99)00007-9

CrossRef Full Text | Google Scholar

Mienert, J., Bünz, S., Guidard, S., Vanneste, M., and Berndt, C. (2005). Ocean bottom seismometer investigations in the ormen lange area offshore mid-Norway provide evidence for shallow gas layers in subsurface sediments. Mar. Petroleum Geol. 22 (1–2), 287–297. doi:10.1016/j.marpetgeo.2004.10.020

CrossRef Full Text | Google Scholar

Reasnor, M., Beaudoin, G., Pfister, M., Ahmed, I., Davis, S., Roberts, M., et al. (2010). “Atlantis time-lapse ocean bottom node Survey: a Project team’s journey from acquisition through processing,” in SEG international exposition and annual meeting (Houston, TX: Society of Exploration Geophysicists). Available online at: https://onepetro.org/SEGAM/proceedings-abstract/SEG10/All-SEG10/96850.

Google Scholar

Ronholt, G. A., Aronsen, H., Guttormsen M, S., Johansen, S., and Klefstad, L. (2008). “Improved imaging using Ocean bottom seismic in the snøhvit field,” in 70th EAGE Conference and exhibition incorporating SPE EUROPEC 2008 (Rome, Italy: European Association of Geoscientists & Engineers). Available online at: https://www.earthdoc.org/content/papers/10.3997/2214-4609.20147591.

Google Scholar

Shakir, U., Ali, A., Hussain, M., Radwan, A. E., and Aal, A. A. E. (2024). PNN enhanced seismic inversion for porosity modeling and delineating the potential heterogeneous gas sands via comparative inversion analysis in the lower indus basin. Pure Appl. Geophys. 181 (9), 2801–2821. doi:10.1007/s00024-024-03562-5

CrossRef Full Text | Google Scholar

Smit, F., Perkins, C., Lepre, L., Craft, K., and Woodard, R. (2008). “Seismic data acquisition using ocean bottom seismic nodes at the deimos field, Gulf of Mexico,” in SEG technical program expanded abstracts (Houston, TX: Society of Exploration Geophysicists), 998–1002. Available online at: https://library.seg.org/doi/abs/10.1190/1.3063805.

CrossRef Full Text | Google Scholar

Stone, J., Wolfarth, S., Prastowo, H., Priyambodo, D., Manning, T., and Etgen, J. (2018). “Tangguh ISS® ocean-bottom node program: a step change in data density, cost efficiency, and image quality,” in SEG international exposition and annual meeting (Houston, TX: Society of Exploration Geophysicists). Available online at: https://onepetro.org/SEGAM/proceedings-abstract/SEG18/All-SEG18/103505.

Google Scholar

Sutton, K., and Gnoffo, P. (1998). “Multi-component diffusion with application to computational aerothermodynamics,” in 7th AIAA/ASME joint thermophysics and heat transfer conference (Albuquerque, NM, U.S.A.: American Institute of Aeronautics and Astronautics). Available online at: https://arc.aiaa.org/doi/10.2514/6.1998-2575.

Google Scholar

Sun, S. Z., Wang, H., Liu, Z., Li, Y., Zhou, X., and Wang, Z. (2012). The theory and application of DEM-gassmann rock physics model for complex carbonate reservoirs. Lead. Edge 31 (2), 152–158. doi:10.1190/1.3686912

CrossRef Full Text | Google Scholar

Tellier, N., and Herrmann, P. (2023). MEMS-based OBN: lessons learnt from the largest OBN survey worldwide. First Break 41 (11), 71–79. doi:10.3997/1365-2397.fb2023093

CrossRef Full Text | Google Scholar

Valdez, A. R., Moreira, P. H. S., Drexler, S., and Couto, P. (2025). A good fit and a better fit. What can relative permeabilities tell us about the Brazilian pre-salt? Geoenergy Sci. Eng. 246, 213567. doi:10.1016/j.geoen.2024.213567

CrossRef Full Text | Google Scholar

Veeken, P. C. H., and Rauch-Davies, M. (2006). AVO attribute analysis and seismic reservoir characterization. First Break 24 (2). doi:10.3997/1365-2397.2006004

CrossRef Full Text | Google Scholar

Wu, Z. Q., Zhang, X. H., Zhao, W. N., Qi, J. H., Zhu, X. Q., and Tian, Z. X. (2021). Ocean Bottom Station Nodes(OBN):progress and achievement. Prog. Geophys. 36 (1), 412–424. doi:10.6038/pg2021EE0029

CrossRef Full Text | Google Scholar

Wu, H., Greer, S. Y., and O’Malley, D. (2023). Physics-embedded inverse analysis with algorithmic differentiation for the earth’s subsurface. Sci. Rep. 13 (1), 718. doi:10.1038/s41598-022-26898-1

PubMed Abstract | CrossRef Full Text | Google Scholar

Xu, A., and Heagy, L. J. (2025). Toward understanding the benefits of neural network parameterizations in geophysical inversions: a study with neural fields. IEEE Trans. Geosci. Remote Sens. 63, 1–14. doi:10.1109/tgrs.2025.3583970

CrossRef Full Text | Google Scholar

Xue, Z. G., Zhang, Z. B., Liu, Z., and Qiu, N. G. (2024). Application of OBN with wide azimuth seismic survey in buried hill exploration. South China J. Seismol. 44 (2), 102–109. doi:10.13512/j.hndz.2024.02.12

CrossRef Full Text | Google Scholar

Yang, T., Liu, Y., and Yang, J. (2023). Joint towed streamer and ocean-bottom-seismometer data multi-parameter full waveform inversion in acoustic-elastic coupled media. Front. Earth Sci. 10, 1085441. doi:10.3389/feart.2022.1085441

CrossRef Full Text | Google Scholar

Ying, X. Y., and Liu, X. X. (2016). Research status and progress of the seismic rock-physics modeling methods. Geophys. Prospect. Petroleum 55 (3), 309–325. doi:10.3969/j.issn.1000-1441.2016.03.001

CrossRef Full Text | Google Scholar

You, J., Cao, J., Wang, X., and Liu, W. (2021). Shear wave velocity prediction based on LSTM and its application for morphology identification and saturation inversion of gas hydrate. J. Petroleum Sci. Eng. 205, 109027. doi:10.1016/j.petrol.2021.109027

CrossRef Full Text | Google Scholar

Zeng, H., Su, Q., Meng, H., Liu, H., Li, H., and Wang, D. (2023). Wide azimuth seismic data processing technology and application: a case study of tight gas reservoirs in western China. Front. Earth Sci. 11, 1267784. doi:10.3389/feart.2023.1267784

CrossRef Full Text | Google Scholar

Zerafa, C., Galea, P., and Sebu, C. (2025). Synergizing deep learning and full-waveform inversion: bridging data-driven and theory-guided approaches for enhanced seismic imaging. arXiv. doi:10.48550/arXiv.2502.17585

CrossRef Full Text | Google Scholar

Zhang, Z. B., and Luo, W. (2021). Analysis of wide orientation of tow cable and effect of both sides. Mar. Geol. Front. 37 (3), 66–73. doi:10.16028/j.1009-2722.2020.037

CrossRef Full Text | Google Scholar

Zhang, J., Yang, Q. L., and Wang, T. Q. (2010). Probing for seismic inversion in depth domain. Oil Geophys. Prospect. 45 (S1), 114–116. doi:10.13810/j.cnki.issn.1000-7210.2010.s1.018

CrossRef Full Text | Google Scholar

Zhang, M. G., Luo, F., Wei, G. W., Ding, G. D., and Chen, Y. S. (2019). Planning and integration of acquisition technologies for ultralarge seismic exploration project, the United Arab Emirates. Nat. Gas Explor. Dev. 42 (2), 66–75.

Google Scholar

Zhang, D., Tsingas, C., Ghamdi, A. A., Huang, M., and Jeong, W. (2020). “Practical issues and solutions of OBN processing,” in SEG technical program expanded abstracts 2020 (Houston, TX: Society of Exploration Geophysicists), 3314–3318. Available online at: https://library.seg.org/doi/10.1190/segam2020-3424895.1.

Google Scholar

Zhang, W., Gao, J., Gao, Z., and Chen, H. (2021). Adjoint-driven deep-learning seismic full-waveform inversion. IEEE Trans. Geosci. Remote Sens. 59 (10), 8913–8932. doi:10.1109/tgrs.2020.3044065

CrossRef Full Text | Google Scholar

Zhang, Y., Liu, Y., Yi, J., and Liu, X. (2021). Fast least-squares reverse time migration of OBN down-going multiples. Front. Earth Sci. 9, 730476. doi:10.3389/feart.2021.730476

CrossRef Full Text | Google Scholar

Zhou, X., Ba, J., Santos, J. E., Carcione, J. M., Fu, L. Y., and Pang, M. (2021). Fluid discrimination in ultra-deep reservoirs based on a double double-porosity theory. Front. Earth Sci. 9, 649984. doi:10.3389/feart.2021.649984

CrossRef Full Text | Google Scholar