Scholarlink

<p>In this work, a hybrid solver with deep learning is proposed for numerical modeling of fluid flow in porous media. The classical simulation procedure is complemented with a neural network model to obtain an initial guess for fluid saturation that is closer to the solution in the Newton–Raphson iterative algorithm. The simulation setup is a 3-dimensional immiscible two-phase flow with fluid motion caused by multiple production and injection wells. Approximation of the initial guess with a neural network model accelerates the numerical modeling up to 14% in terms of nonlinear iterations. Extensive experiments with dynamic and static reservoir features revealed that improving predictive accuracy does not necessarily improve fluid modeling. Different training procedures (e.g., different loss functions) and feature spaces (e.g., more past time steps used) can lead to better prediction quality but a higher number of nonlinear iterations. These results demonstrate that not only the closeness to the solution, but also the spatial distribution of residuals affects the “quality” of the starting point in Newton’s method.</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

Hybrid solver with deep learning for transport problem in porous media

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月31日

查看全文
  • Vladislav Trifonov,
  • Egor Illarionov,
  • Anton Voskresenskii,
  • Musheg Petrosyants,
  • Klemens Katterbauer

摘要

In this work, a hybrid solver with deep learning is proposed for numerical modeling of fluid flow in porous media. The classical simulation procedure is complemented with a neural network model to obtain an initial guess for fluid saturation that is closer to the solution in the Newton–Raphson iterative algorithm. The simulation setup is a 3-dimensional immiscible two-phase flow with fluid motion caused by multiple production and injection wells. Approximation of the initial guess with a neural network model accelerates the numerical modeling up to 14% in terms of nonlinear iterations. Extensive experiments with dynamic and static reservoir features revealed that improving predictive accuracy does not necessarily improve fluid modeling. Different training procedures (e.g., different loss functions) and feature spaces (e.g., more past time steps used) can lead to better prediction quality but a higher number of nonlinear iterations. These results demonstrate that not only the closeness to the solution, but also the spatial distribution of residuals affects the “quality” of the starting point in Newton’s method.