Scholarlink

<p>Deep learning has made much progress in seismic impedance inversion. Popular deep architectures employ convolution layers and fully connected layers. However, they often suffer from high learning cost, insufficient lateral continuity, and huge training data requirements. In this paper, we propose a new approach combining residual block and Kolmogorov–Arnold network (KAN) to address these issues. First, the original convolutions in residual blocks and the fully connected layers are replaced by dilated convolutions and KAN, respectively. This technique improves inversion accuracy without significantly increasing the number of parameters. Second, the contour features are extracted to participate in the calculation of the loss function. This constraint helps to enhance lateral continuity and alleviate network overfitting. Third, a transfer learning technique is designed to improve the generalization of our network. It mitigates the challenges associated with training deep learning networks from scratch on small datasets. Experiments are undertaken on synthetic seismic dataset from Marmousi2 and Overthrust models. The results show that the new approach outperforms other data-driven approaches in terms of accuracy, robustness, and generalization. The source code is available at <a href="https://github.com/x3nny/reskan">https://github.com/x3nny/reskan</a>.</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

ResKAN: residual Kolmogorov–Arnold network with transfer learning for seismic impedance inversion

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Tian-Yi Luo,
  • Jin-Yu Tang,
  • Shulin Pan,
  • Ziyu Qin,
  • Fan Min

摘要

Deep learning has made much progress in seismic impedance inversion. Popular deep architectures employ convolution layers and fully connected layers. However, they often suffer from high learning cost, insufficient lateral continuity, and huge training data requirements. In this paper, we propose a new approach combining residual block and Kolmogorov–Arnold network (KAN) to address these issues. First, the original convolutions in residual blocks and the fully connected layers are replaced by dilated convolutions and KAN, respectively. This technique improves inversion accuracy without significantly increasing the number of parameters. Second, the contour features are extracted to participate in the calculation of the loss function. This constraint helps to enhance lateral continuity and alleviate network overfitting. Third, a transfer learning technique is designed to improve the generalization of our network. It mitigates the challenges associated with training deep learning networks from scratch on small datasets. Experiments are undertaken on synthetic seismic dataset from Marmousi2 and Overthrust models. The results show that the new approach outperforms other data-driven approaches in terms of accuracy, robustness, and generalization. The source code is available at https://github.com/x3nny/reskan.