还原(数学)
近似误差
替代模型
均方误差
块(置换群论)
算法
推论
相关系数
压力(语言学)
相似性(几何)
曲面(拓扑)
压力测试(软件)
计算机科学
数据挖掘
计算机模拟
模拟
计算模型
统计
平均绝对误差
作者
Wenlong Xia,Hui Zhao,Zhaoyu Wang,Fankun Meng,Bin Shen,Tianyu Yang
出处
期刊:Physics of Fluids
[American Institute of Physics]
日期:2026-11-01
卷期号:38 (11): 117408-
摘要
Geological carbon dioxide (CO2) storage involves coupled fluid-flow and geomechanical processes and is a key technology for carbon mitigation. During storage, pore-pressure perturbations induced by CO2 injection propagate through the reservoir and surrounding formations, alter the effective stress field, and may trigger geomechanical responses such as surface deformation. Conventional hydro-mechanical numerical simulation is computationally expensive and therefore difficult to use for rapid risk assessment. This study proposes a surrogate model that combines bidirectional convolutional long short-term memory (BiConvLSTM), a U-Net architecture, and a multi-scale adaptive spatial attention gate (MASAG), hereafter referred to as BiConvLSTM-UNet-MASAG, together with physics-constrained loss terms to jointly predict the three-dimensional pore-pressure field, the three-dimensional average effective stress field, and surface vertical displacement. A gridded hydro-mechanical model was established using the Generalized Equation-of-State Modeling simulator developed by Computer Modelling Group Ltd. (CMG), and 1000 samples were generated by varying porosity, permeability, and injection rate. On the testing set, the proposed surrogate model achieved field-specific coefficient of determination (R2) values of 0.98, 0.97, and 0.98 for pore pressure, average effective stress, and surface vertical displacement, respectively. The corresponding structural similarity index values for the three-dimensional pore-pressure and effective-stress fields were 0.92 and 0.91. Field-specific comparisons further showed that BiConvLSTM-UNet-MASAG consistently reduced root mean square error and mean absolute error relative to both the baseline BiConvLSTM-UNet and the variant enhanced with a convolutional block attention module across the three prediction tasks. The inference time for one case was 1.3 s, whereas a CMG simulation required 726 min. The model balances multi-field prediction accuracy and computational efficiency while efficiently reproducing coupled pressure, stress, and deformation responses, supporting rapid assessment and monitoring interpretation for geological CO2 storage.
科研通智能强力驱动
Strongly Powered by AbleSci AI