Convolutional autoencoder network lithology recognition based on scratch tests.

Area of Science:

• Geology
• Artificial Intelligence
• Petroleum Engineering

Background:

• Continental shale formations, such as those in the Songliao Basin, exhibit frequent lithological alternations.
• Accurate lithology identification is crucial for refined reservoir modeling, especially in heterogeneous shale reservoirs.

Purpose of the Study:

• To develop a high-precision lithology identification method for continental shale using mechanical property data.
• To evaluate the performance of a novel convolutional neural network (CNN) and autoencoder network (AE) integrated approach.

Main Methods:

• Scratch tests were performed on shale reservoir cores from the Qingshankou Formation (2360m-2409m).
• Nine mechanical characteristic parameters were obtained, including hardness, compressive strength, and Poisson's ratio.
• A lithology identification method was developed by integrating CNN and AE, with optimal scale selection and comparative analysis against other neural networks.

Main Results:

• The proposed method, using a convolutional autoencoder network (CAE), achieved 89.58% accuracy on the test dataset at an identification scale of 20×9.
• Recall rates exceeded 84% across all lithology recognitions.
• The CAE method demonstrated superior accuracy and recall rates compared to other neural network approaches.

Conclusions:

• The developed CAE-based method offers a novel and precise approach for reservoir lithology identification in complex shale formations.
• This methodology provides a foundation for more accurate modeling of fracture propagation in heterogeneous shale reservoirs.
• The findings support improved reservoir characterization and management in basins with similar geological characteristics.