Identification and Intelligent Prediction of Microscopic Residual Oil Distribution Based on the TransUNet Neural Network.

The morphological distribution and classification of remaining oil are of great significance to oilfield development. Traditional identification and classification methods are limited by large manual errors, low intelligence, and insufficient precision, and fail to accurately identify microscale remaining oil below 10 μm. This paper proposes a deep learning-based method for remaining oil recognition and classification using TransUNet as the backbone network, combined with image augmentation and Transformer multihead attention mechanism to enhance feature extraction and classification performance. Results show that the proposed method achieves an overall classification accuracy of 94%, with effective recognition thresholds of 3.09 μm for film-like remaining oil and 1.54 μm for drip-like remaining oil. In addition, different displacement mechanisms result in significantly different occurrence states and dynamic evolution laws of remaining oil: water flooding relies on mechanical scouring, polymer flooding depends on viscosity improvement and profile control, while surfactant flooding achieves efficient displacement by reducing interfacial tension and generating emulsification. Among them, remaining oil formed by surfactant flooding is mainly droplet-shaped and easy to be displaced, which reveals its superior displacement effect from a microscopic perspective and provides a basis for formulating targeted oilfield development measures.