Scholarlink

<p>As a core technology in seismic exploration, full-waveform inversion (FWI) realizes the inversion from seismic waveform data to underground media parameters by solving the nonlinear optimization problem driven by wave equations. Traditional FWI relies heavily on the initial model, while deep learning methods have the problems of insufficient resolution and smoothed boundaries. In order to solve this problem, a high-precision inversion method based on multi-scale dual-encoder diffusion model was proposed. The cosine modulation noise scheduling strategy is used to optimize the reverse diffusion process by combining the implicit model of denoising diffusion. Using the asymmetric encoder--decoder structure, the AttInverNet network is trained to extract the multi-scale features of seismic data, which is integrated into the noise predictor to achieve feature dimensionality reduction and key information retention, and the cross-attention mechanism is integrated to construct a seismic data-driven velocity imaging model. Experimental results show that compared with the InversionNet and MAU-net methods, the proposed model has the best performance in loss convergence and evaluation indexes, and can more accurately reconstruct the complex fault structure and velocity distribution, which provides an effective solution for complex geological inversion.</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

Seismic full-waveform inversion based on multi-scale dual-encoding diffusion model

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Yongming Huang,
  • Pan Wang,
  • Bo Xue,
  • Yejun Sun,
  • Yong Lu

摘要

As a core technology in seismic exploration, full-waveform inversion (FWI) realizes the inversion from seismic waveform data to underground media parameters by solving the nonlinear optimization problem driven by wave equations. Traditional FWI relies heavily on the initial model, while deep learning methods have the problems of insufficient resolution and smoothed boundaries. In order to solve this problem, a high-precision inversion method based on multi-scale dual-encoder diffusion model was proposed. The cosine modulation noise scheduling strategy is used to optimize the reverse diffusion process by combining the implicit model of denoising diffusion. Using the asymmetric encoder--decoder structure, the AttInverNet network is trained to extract the multi-scale features of seismic data, which is integrated into the noise predictor to achieve feature dimensionality reduction and key information retention, and the cross-attention mechanism is integrated to construct a seismic data-driven velocity imaging model. Experimental results show that compared with the InversionNet and MAU-net methods, the proposed model has the best performance in loss convergence and evaluation indexes, and can more accurately reconstruct the complex fault structure and velocity distribution, which provides an effective solution for complex geological inversion.