Scholarlink

<p>Well log reconstruction endeavors to restore incomplete or corrupted data in well log records. Accurate reconstruction of well log data furnishes a robust and reliable foundation for lithofacies discrimination, reservoir delineation, hydrocarbon appraisal, and stratigraphic subdivision. In recent years, data-driven methodologies for well log reconstruction have witnessed remarkable progress. To satisfy the extensive data demands intrinsic to deep learning paradigms, contemporary research predominantly integrates intra-field datasets for model training. However, this approach often overlooks the intrinsic geological context and interpretive underpinnings embedded in well log response patterns, thus constraining further improvements in model efficacy. To address this limitation, we introduce a novel well log reconstruction framework, designated as Diff-Mamba. Specifically, our method synergistically incorporates the selective state space model (SSSM)—which implements a linearly scalable attention paradigm—with diffusion-based generative models (denoising diffusion probabilistic models, DDPM), thereby optimizing training efficiency and model expressiveness. Moreover, leveraging the dynamically adaptive step size within the selective state space enables the model to autonomously adapt to missing depth intervals during training. Furthermore, the decoding stage employs modal decomposition to perform multiscale frequency analysis on well log data, allowing the original signal to be partitioned into spectral components that encapsulate geological features across disparate scales. This facilitates efficient and interpretable data augmentation for scarce lithological horizons. Experimental results indicate that, following data augmentation, our method substantially improves the reconstruction accuracy of deep resistivity logs (RMSE: 0.4311, MAE: 0.3243, R<sup>2</sup>: 84.52%) compared to the baseline (RMSE: 0.5079, MAE: 0.3620, R<sup>2</sup>: 80.34%). For the reconstruction of acoustic logs and cross-regional resistivity logs, augmenting the training set with 2,000 synthetic samples resulted in R<sup>2</sup> values of 90.02% and 85.69%, respectively, further confirming the efficacy and cross-domain generalizability of the proposed framework.</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

Geophysical Logging Curve Reconstruction Based on Diff-Mamba Modeling

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Guozeng Zhang,
  • Tan Chen,
  • Jialin Chen,
  • Junkai Chen,
  • Cheng Zhang,
  • Ziyan Feng,
  • Cheng Feng

摘要

Well log reconstruction endeavors to restore incomplete or corrupted data in well log records. Accurate reconstruction of well log data furnishes a robust and reliable foundation for lithofacies discrimination, reservoir delineation, hydrocarbon appraisal, and stratigraphic subdivision. In recent years, data-driven methodologies for well log reconstruction have witnessed remarkable progress. To satisfy the extensive data demands intrinsic to deep learning paradigms, contemporary research predominantly integrates intra-field datasets for model training. However, this approach often overlooks the intrinsic geological context and interpretive underpinnings embedded in well log response patterns, thus constraining further improvements in model efficacy. To address this limitation, we introduce a novel well log reconstruction framework, designated as Diff-Mamba. Specifically, our method synergistically incorporates the selective state space model (SSSM)—which implements a linearly scalable attention paradigm—with diffusion-based generative models (denoising diffusion probabilistic models, DDPM), thereby optimizing training efficiency and model expressiveness. Moreover, leveraging the dynamically adaptive step size within the selective state space enables the model to autonomously adapt to missing depth intervals during training. Furthermore, the decoding stage employs modal decomposition to perform multiscale frequency analysis on well log data, allowing the original signal to be partitioned into spectral components that encapsulate geological features across disparate scales. This facilitates efficient and interpretable data augmentation for scarce lithological horizons. Experimental results indicate that, following data augmentation, our method substantially improves the reconstruction accuracy of deep resistivity logs (RMSE: 0.4311, MAE: 0.3243, R2: 84.52%) compared to the baseline (RMSE: 0.5079, MAE: 0.3620, R2: 80.34%). For the reconstruction of acoustic logs and cross-regional resistivity logs, augmenting the training set with 2,000 synthetic samples resulted in R2 values of 90.02% and 85.69%, respectively, further confirming the efficacy and cross-domain generalizability of the proposed framework.