Lithofacies identification constitutes a core component of reservoir characterization and is of great significance for oil and gas exploration and development. Traditional lithofacies identification relies primarily on expert experience and manual interpretation, suffering from high subjectivity, low efficiency, and poor consistency. Moreover, identifying carbonate lithofacies such as bioreefs and bioclastic shoals solely through geological and geophysical data combined with limited core samples and thin-section analysis is labor-intensive and cannot meet the requirements of refined oil and gas field development. This study aims to address the major challenges in intelligent lithofacies identification, namely the insufficient modeling of spatial dependencies in the depth direction, the lack of technical prior knowledge, and the inadequate integration of geological understanding with data-driven approaches.
We propose an intelligent well-logging lithofacies identification method for deep carbonate rocks that integrates sequence stratigraphic prior information. The method applies a deep neural network, termed Sequence-Constrained Long Short-Term Memory Network (SC-LSTM), as its core framework. This network embeds sequence stratigraphic units as geological constraints into the input feature space of the model, achieving an organic integration of geological understanding with logging response characteristics. The network architecture comprises two parallel feature extraction modules: a curve feature extraction module that utilizes convolutional layers to extract local patterns and depth-directional variation characteristics from logging responses, and a sequence feature extraction module that performs convolutional encoding on sequence division schemes to convert discrete sequence and system tract information into continuous feature representations. The method uses conventional well logs, including natural gamma ray, acoustic transit time, density, neutron, and resistivity as input data. A sliding time-window sampling strategy expands the training dataset to address the problem of limited labeled samples, while a weighted averaging prediction method enhances the stability and reliability of the identification results.
The application to the Changxing Formation carbonate reservoir in the Yuanba gas field demonstrates that an optimal sample length of 40 meters can (i) completely cover the vertical evolution sequence from reef base to reef core to reef cap, or (ii) encompass a complete reef-shoal depositional combination controlled by sea-level cycles. This, enables the model to fully learn the orderly stacking patterns and facies transition characteristics within sequence units. The model achieved 92% identification accuracy on the test set. Comparative experiments between conventional LSTM and SC-LSTM networks reveal that the proposed method effectively avoids geologically unreasonable predictions that may arise from relying solely on logging response characteristics. For example, it avoids erroneously predicting reef cap facies during the early highstand system tract or misidentifying reef cap facies as reef/shoal interfacies at reef tops.
These findings confirm that the integration of sequence stratigraphic prior information can effectively constrain model predictions and avoid geologically unreasonable results. The conventional LSTM model learns lithofacies characteristics only from numerical patterns of logging curves and easily confuses lithofacies types with similar logging responses under different sequence backgrounds. In contrast, the SC-LSTM model enables the network to understand the geological rules and the combined characteristics of lithofacies development within different sequence units and system tracts by fusing sequence stratigraphic information. This improves the geological rationality and prediction accuracy of lithofacies identification. [Significance] The significance of this study lies not only in its practical applications—enabling rapid establishment of regional lithofacies distribution patterns to guide well deployment and providing timely support for reservoir fine characterization and the exploitation of remaining oil and gas potential through the real-time updating of lithofacies interpretation schemes as new well data accumulate—but also in its theoretical contributions. It particularly advances the integration of domain knowledge with artificial intelligence technology by linking sequence stratigraphy theory, depositional environment evolution, and deep learning algorithms in a unified framework for intelligent lithofacies identification. This opens new avenues for automated and intelligent reservoir characterization in complex carbonate formations.
| 科 Family | 属数 Number of genus | 种数 Number of species | 占总种数比例 Percentage of total species (%) | 属 Genus | 种数 Number of species | 占总种数比例 Percentage of total species (%) |
|---|---|---|---|---|---|---|
| 鹅膏菌科Amanitaceae | 2 | 11 | 5.26 | 鹅膏菌属 Amanita | 10 | 4.78 |
| 小菇科 Mycenaceae | 2 | 12 | 5.74 | 丝盖伞属 Inocybe | 5 | 2.39 |
| 多孔菌科 Polyporaceae | 8 | 14 | 6.70 | 蜡蘑属 Laccaria | 5 | 2.39 |
| 红菇科 Russulaceae | 3 | 23 | 11.00 | 小皮伞属 Marasmius | 6 | 2.87 |
| 小菇属 Mycena | 11 | 5.26 | ||||
| 光柄菇属 Pluteus | 5 | 2.39 | ||||
| 红菇属 Russula | 17 | 8.13 | ||||
| 栓菌属 Trametes | 5 | 2.39 |