Scholarlink

<p>Seismic inversion that estimates subsurface rock elastic parameters from observed seismic data faces great challenges from the non-uniqueness of the solutions. Many efforts have been dedicated to improving the stability of inverse problem solutions by incorporating regularization techniques based on prior information, yet achieving high inversion accuracy and resolution for the subsurface parameters remains challenging. Recently, deep learning (DL) has emerged as a promising alternative approach for solving complex inversion problems. However, the incorporation of DL in seismic inversion faces many challenges, for example, the requirement for labeled data for training, and the occasional generation of predictions that are difficult to interpret. To address these problems, an independent new seismic impedance inversion method based on self-supervised learning (without the need for labels) is proposed herein. This method utilizes a temporal convolutional network to extract temporal features from seismic data and incorporates a Robinson convolutional model to ensure that the inversion outputs adhere to the established physical laws. Furthermore, the proposed method introduces a sparsity constraint layer to improve solution sparsity and incorporates prior information in the loss function to mitigate the non-uniqueness of seismic inversion. Together, these components establish a closed-loop, self-supervised network framework that trains the model parameters by iterating the discrepancy between observed seismic data and the model’s predictions. The application to a synthetic seismic dataset suggests that the proposed method outperforms traditional methods (e.g., smooth-constrained and sparsity-constrained inversion methods) in terms of inversion accuracy. Also, the self-learning scheme is superior to semi-supervised methods because the former can achieve high accuracy and does not require labels. This superior performance of the physical-informed DL method is further demonstrated in a real-world application, notably for generating higher-resolution images of subsurface seismic impedance.</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

Physics-Guided Self-Supervised Seismic Impedance Inversion

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Detao Wang,
  • Guoxiong Chen,
  • Qiuming Cheng

摘要

Seismic inversion that estimates subsurface rock elastic parameters from observed seismic data faces great challenges from the non-uniqueness of the solutions. Many efforts have been dedicated to improving the stability of inverse problem solutions by incorporating regularization techniques based on prior information, yet achieving high inversion accuracy and resolution for the subsurface parameters remains challenging. Recently, deep learning (DL) has emerged as a promising alternative approach for solving complex inversion problems. However, the incorporation of DL in seismic inversion faces many challenges, for example, the requirement for labeled data for training, and the occasional generation of predictions that are difficult to interpret. To address these problems, an independent new seismic impedance inversion method based on self-supervised learning (without the need for labels) is proposed herein. This method utilizes a temporal convolutional network to extract temporal features from seismic data and incorporates a Robinson convolutional model to ensure that the inversion outputs adhere to the established physical laws. Furthermore, the proposed method introduces a sparsity constraint layer to improve solution sparsity and incorporates prior information in the loss function to mitigate the non-uniqueness of seismic inversion. Together, these components establish a closed-loop, self-supervised network framework that trains the model parameters by iterating the discrepancy between observed seismic data and the model’s predictions. The application to a synthetic seismic dataset suggests that the proposed method outperforms traditional methods (e.g., smooth-constrained and sparsity-constrained inversion methods) in terms of inversion accuracy. Also, the self-learning scheme is superior to semi-supervised methods because the former can achieve high accuracy and does not require labels. This superior performance of the physical-informed DL method is further demonstrated in a real-world application, notably for generating higher-resolution images of subsurface seismic impedance.