Core-based recognition of well proppant particles using an enhanced ResNet model.

Area of Science:

• Petroleum Engineering
• Artificial Intelligence in Geoscience

Background:

• Accurate proppant particle identification in drill cuttings is crucial for understanding hydraulic fracturing behavior and reservoir fracture geometry.
• Manual identification methods are labor-intensive, require specialized expertise, and are prone to subjective bias.

Purpose of the Study:

• To develop an enhanced ResNet-based model for automated proppant particle recognition in drill cuttings.
• To improve the accuracy and efficiency of proppant identification compared to manual methods.

Main Methods:

• Developed an enhanced ResNet model incorporating multi-scale linear deformable convolution (MDS_LDconv), an Adaptive Star operation block, and a Mamba-inspired linear attention module (MLLA_G).
• Replaced the conventional fully connected output layer with a KAN layer for improved processing of complex cutting images.

Main Results:

• The enhanced model achieved a test-set accuracy of 98.62%, outperforming standard ResNet-50 by 2.16%.
• Field trials showed 96.31% agreement between machine identification and manual interpretation.
• Inference time was reduced to only 4% of human analysis time, demonstrating significant efficiency gains.

Conclusions:

• The developed AI model offers a highly accurate and efficient solution for proppant particle identification in hydraulic fracturing.
• This automated approach substantially reduces manpower expenditure and subjective bias, enhancing the reliability of fracture characterization.