Fuel | A new approach to hydraulic fracturing prediction in tight gas reservoir based on clustering enhanced multi-information fusion deep learning model (CMFD)

By  李勇    2026-09-29    Visited 10 times

Abstract

Quantitative characterization and prediction of hydraulic fractures in tight gas reservoirs are the core problems in the effective production from tight gas reservoirs at present. Fracture distribution is complexly influenced by multiple factors, including rock mechanics and reservoir characteristics, and the fracture propagation process exhibits high nonlinearity. Conventional fracture propagation simulation methods not only have extremely high computational costs but also face convergence problems, making the prediction of fracture distribution characteristics under complex conditions a persistent issue in the current development of tight gas reservoirs. Therefore, this research proposes an improved artificial intelligence (AI) approach for hydraulic fracturing prediction using a multi-source fusion deep learning model integrated with clustering information, which was developed by combining unsupervised clustering and multi-source information, including 3D geological-rock mechanical (GM) modeling, seismic, and micro-seismic monitoring data.

The study was carried out in the Shihezi Formation of the Linxing tight-gas field. The basic geological model of the Linxing Block was constructed under the constraint of seismic data using the sequential Gaussian simulation (SGS) method. Subsequently, based on the geological model, the Young’s modulus, Poisson’s ratio, and other engineering mechanics models closely related to fracturing in the study area were constructed. Unsupervised bisecting k-means (bi-k-means) clustering was performed on preprocessed reservoir data and AI prediction accuracy. All 3D GM models, clustering models, and two-dimensional (2D) hydraulic fracture (HF) energy data were used to train the U-Net and develop the multi-information fusion U-Net (MUM) model for predicting HF morphologies and geometric parameters through 3 output channels. The results were validated by micro-seismic monitoring, confirming the physical validity of the model. The 6-layer MUM model achieved high predictive accuracy of approximately 94.9 % using the cross-entropy (CE) loss function and 20 input channels. Results indicate that the bi-k-means (k = 4) fusion with the U-Net algorithm yielded improved prediction.

Paper Information:

TANG Mingming, Bushra Mohsin. A new approach to hydraulic fracturing prediction in tight gas reservoir based on clustering enhanced multi-information fusion deep learning model (CMFD)[J]. Fuel, 2026. Doi: https://doi.org/10.1016/j.fuel.2026.140619.