Scholarlink

<p>Formation pore pressure estimation is pivotal in various aspects of geosciences and the energy industry, influencing geomechanical assessments, energy production, carbon storage, and risk management. Despite decades of research, accurate pore pressure prediction and its changes remain challenging in sedimentary reservoirs, necessitating alternative solutions beyond traditional methods. This study develops models that reliably predict pore pressure changes using six machine learning (ML) algorithms applied to a large laboratory-derived dataset (14,532 data records). Pre-processing was conducted to ensure accuracy and generalizability, including outlier removal. Five stress-related variables were used as input features for the ML models: vertical and horizontal effective stresses, shear stress, mean stress, and octahedral shear stress. Tailored algorithms were applied to the dataset to train, test, and evaluate the performance of the ML models. The Gaussian process regression (GPR) model exhibited the lowest RMSE during both training (0.00085&#xa0;MPa) and testing (0.00209&#xa0;MPa) stages. Robustness analysis highlighted GPR’s high tolerance threshold to noisy data. SHAP analysis revealed the significant impact of the vertical stress on the GPR model’s prediction, whereas shear stress exerted the least impact on the pore pressure predictions. Partial dependency analysis identifies highly complex relationships between the input variables and pore pressure. The proposed framework is calibrated on isothermal (~ 22&#xa0;°C) laboratory hydrostatic-test data and is therefore positioned as a stress-to-pore-pressure surrogate and a proof of concept; thermal effects, poroelastic coupling, fluid migration, and validation against field injection and downhole-monitoring data are identified as essential next steps before operational deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine-learning modeling of pore pressure variation in geological CO2 storage reservoir

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Shadfar Davoodi,
  • Hung Vo Thanh,
  • David A. Wood,
  • Mohammad Mehrad,
  • Evgeny Burnaev

摘要

Formation pore pressure estimation is pivotal in various aspects of geosciences and the energy industry, influencing geomechanical assessments, energy production, carbon storage, and risk management. Despite decades of research, accurate pore pressure prediction and its changes remain challenging in sedimentary reservoirs, necessitating alternative solutions beyond traditional methods. This study develops models that reliably predict pore pressure changes using six machine learning (ML) algorithms applied to a large laboratory-derived dataset (14,532 data records). Pre-processing was conducted to ensure accuracy and generalizability, including outlier removal. Five stress-related variables were used as input features for the ML models: vertical and horizontal effective stresses, shear stress, mean stress, and octahedral shear stress. Tailored algorithms were applied to the dataset to train, test, and evaluate the performance of the ML models. The Gaussian process regression (GPR) model exhibited the lowest RMSE during both training (0.00085 MPa) and testing (0.00209 MPa) stages. Robustness analysis highlighted GPR’s high tolerance threshold to noisy data. SHAP analysis revealed the significant impact of the vertical stress on the GPR model’s prediction, whereas shear stress exerted the least impact on the pore pressure predictions. Partial dependency analysis identifies highly complex relationships between the input variables and pore pressure. The proposed framework is calibrated on isothermal (~ 22 °C) laboratory hydrostatic-test data and is therefore positioned as a stress-to-pore-pressure surrogate and a proof of concept; thermal effects, poroelastic coupling, fluid migration, and validation against field injection and downhole-monitoring data are identified as essential next steps before operational deployment.