Scholarlink

<p>Porous media reconstruction is a valuable and cost-efficient way to discover the internal microstructure of porous media, and thus helps to reveal the general rules of fluid flow in porous media and complex properties of porous materials. Recently, generative adversarial network (GAN) widely used in deep learning generation constantly promotes the development of porous media. Because GAN has powerful feature extraction capabilities, it can help to generate porous media samples with the characteristics of real porous microstructures. However, the training process of GAN is often affected by gradient issues (e.g., gradient vanishing or gradient explosion). This study introduces a normalization method, namely gradient normalization (GN), to address the gradient issues that arise during GAN’s training for the reconstruction of porous media. Unlike traditional methods, the GN method imposes constraints on the gradient norm of the discriminator function. This approach enhances the model stability and thus increases the network capacity. We assess the efficacy of the GN method in the complex 3D stochastic reconstruction of porous media through a series of comparative experiments, through which GN is proved to be significantly superior to some other reconstruction methods in multiple evaluation metrics, demonstrating the effectiveness and potential of GN in 3D stochastic reconstruction of porous media.</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

Modeling stochastic porous media using gradient normalization based generative adversarial network

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Ting Zhang,
  • Xiangyu Chen,
  • Mengkai Yin,
  • Yuqi Wu,
  • Yi Du

摘要

Porous media reconstruction is a valuable and cost-efficient way to discover the internal microstructure of porous media, and thus helps to reveal the general rules of fluid flow in porous media and complex properties of porous materials. Recently, generative adversarial network (GAN) widely used in deep learning generation constantly promotes the development of porous media. Because GAN has powerful feature extraction capabilities, it can help to generate porous media samples with the characteristics of real porous microstructures. However, the training process of GAN is often affected by gradient issues (e.g., gradient vanishing or gradient explosion). This study introduces a normalization method, namely gradient normalization (GN), to address the gradient issues that arise during GAN’s training for the reconstruction of porous media. Unlike traditional methods, the GN method imposes constraints on the gradient norm of the discriminator function. This approach enhances the model stability and thus increases the network capacity. We assess the efficacy of the GN method in the complex 3D stochastic reconstruction of porous media through a series of comparative experiments, through which GN is proved to be significantly superior to some other reconstruction methods in multiple evaluation metrics, demonstrating the effectiveness and potential of GN in 3D stochastic reconstruction of porous media.