Scholarlink

<p>Geostatistical techniques are widely used for modeling reservoir heterogeneity and assessing uncertainty. Although multiple geostatistical reservoir models are typically generated, due to computational costs, only a small fraction can be utilized for comprehensive flow simulation. In this study, we propose an efficient data-driven framework to identify reservoir connectivity, which can be used as a criterion for screening and ranking geostatistical reservoir models. While pure data-driven models are widely used for their efficiency and simplicity, they often lack physical interpretability and predictive power. To address these challenges, a physics-informed machine learning framework is employed. Our machine learning framework takes routine well measurements, including injection rate and pressure data, as input, and multiphase production rates as output. To enhance predictive power while maintaining efficiency, a reduced-physics model is incorporated into a physics-informed neural network (PINN) architecture. Physics-based regularization terms are embedded in the loss function to incorporate the governing partial differential equations (PDEs) of the reduced-physics system. The trained PINN model satisfies these governing PDEs, making it physically interpretable and capable of identifying preferential fluid flow paths in terms of flux allocation. The estimated flux allocation helps screen and rank geostatistical reservoir models, significantly reducing computational burden compared to full-physics flow simulations. Furthermore, the physics-based regularization enables the PINN model to avoid overfitting and improves its predictive performance. The power and efficacy of our proposed framework are demonstrated using a benchmark reservoir simulation case and a field water-flooding case. The proposed PINN method delivers forecasting performance for multiphase production rates with reasonable accuracy. The flux allocations estimated by the trained PINN model show good agreement with those obtained from streamline-based methods. Moreover, the proposed screening workflow successfully ranks and selects plausible geostatistical reservoir models in a computationally efficient manner.</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

Screening Geostatistical Models Through Identification of Preferential Flow Paths Using Physics Informed Machine Learning

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Masahiro Nagao,
  • Sauvik Das,
  • Akhil Datta-Gupta

摘要

Geostatistical techniques are widely used for modeling reservoir heterogeneity and assessing uncertainty. Although multiple geostatistical reservoir models are typically generated, due to computational costs, only a small fraction can be utilized for comprehensive flow simulation. In this study, we propose an efficient data-driven framework to identify reservoir connectivity, which can be used as a criterion for screening and ranking geostatistical reservoir models. While pure data-driven models are widely used for their efficiency and simplicity, they often lack physical interpretability and predictive power. To address these challenges, a physics-informed machine learning framework is employed. Our machine learning framework takes routine well measurements, including injection rate and pressure data, as input, and multiphase production rates as output. To enhance predictive power while maintaining efficiency, a reduced-physics model is incorporated into a physics-informed neural network (PINN) architecture. Physics-based regularization terms are embedded in the loss function to incorporate the governing partial differential equations (PDEs) of the reduced-physics system. The trained PINN model satisfies these governing PDEs, making it physically interpretable and capable of identifying preferential fluid flow paths in terms of flux allocation. The estimated flux allocation helps screen and rank geostatistical reservoir models, significantly reducing computational burden compared to full-physics flow simulations. Furthermore, the physics-based regularization enables the PINN model to avoid overfitting and improves its predictive performance. The power and efficacy of our proposed framework are demonstrated using a benchmark reservoir simulation case and a field water-flooding case. The proposed PINN method delivers forecasting performance for multiphase production rates with reasonable accuracy. The flux allocations estimated by the trained PINN model show good agreement with those obtained from streamline-based methods. Moreover, the proposed screening workflow successfully ranks and selects plausible geostatistical reservoir models in a computationally efficient manner.