Work

Deep Learning-Based Adaptive Characterization Identification of Cement Microstructure in Well Cement

Year: 2026

Type: article

Abstract: Cementing slurry serves as a critical safety barrier for deep well cementing. In the gelation stage, cement slurry often forms microstructural pore channels, significantly increase the risk of wellbore seal failure and gas leakage during cyclic injection and production. To address these challenges, a deep learning-based adaptive characterization identification of cement microstructure is proposed. This study employs environmental scanning electron microscopy (ESEM) and semi-supervised image segmentation to simulate and characterize cement microstructures from a microscopic perspective. Imaging cement microstructures is challenging due to blurred edges, irregular feature patterns, and limited annotated data. To overcome these limitations, we propose a novel image segmentation model named Cement Segmentation Network (CemSNet), which integrates a hybrid adaptive attention mechanism. By combining channel and spatial self-attention, CemSNet effectively captures irregular micropore distributions and enhances edge definition, significantly improving the precision of segmenting fine-edge details and small features. To address the scarcity of labeled training data, a semi-supervised learning strategy is adopted. This approach leverages consistency regularization and pseudo-labeling to utilize unlabeled data, thereby enhancing model robustness. With limited labeled samples, the proposed method achieves a Mean Intersection over Union (MIoU) of 81.46%, a Mean Dice coefficient of 89.28%, and a pixel accuracy of 96.24%, demonstrating high segmentation performance. Furthermore, based on segmentation results and physical modeling, this study reveals that both porosity and permeability decrease significantly over time, with rapid reductions in early stages and slower declines thereafter. Lower water–cement ratios accelerate this reduction, leading to denser microstructures. These findings offer a new perspective for evaluating the sealing performance of wells cement. • Well cement microstructure is analyzed using ESEM. • The CemSNet model with a HAAM module segments ESEM cement images. • A semi-supervised learning framework utilizes unlabeled images. • Combining microstructure and permeability for cement sealing research.

Source: Case Studies in Construction Materials

DOI: 10.1016/j.cscm.2026.e05903

Authors: Xinghao Yang, Zhiyuan Wang, Xuerui Wang, Zhenhao Li, Yonghai Gao, Hao Li

Institution: China University of Petroleum, East China

Language: English

Topic: Drilling and Well Engineering

Subfield: Ocean Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $2,660