Scholarlink

<p>The detection and analysis of the Earth’s free oscillations in the form of normal modes contribute to our understanding of the planet’s structure, dynamics, and seismicity. However, false and dismissed recognitions of the focused normal mode signals prevent the detailed exploration of the Earth’s interior due to the complex patterns of the normal mode spectrum. Here, we develop a deep-learning-based neural network, named ModeNet, which is capable of precisely and efficiently selecting the frequency windows to cover the target normal modes on noisy spectra. ModeNet achieves a remarkable precision rate in the discrimination between normal modes and noises of 98.1%, with very few false positives (normal modes misidentified as noises) and false negatives (noises misidentified as normal modes). ModeNet achieves good generalization in processing seismograms with different noise levels, components, and time windows, as well as superconducting-gravimeter (SG) observations. Extracting the frequencies of normal modes within windows selected by ModeNet and further performing a radial (1D) Earth structure inversion, our preferred model shows a better fit to the observed eigenfrequencies than the Preliminary Reference Earth Model (PREM). Therefore, ModeNet would be implemented as a potentially valuable tool for the future 3D structure inversion of the deep Earth and extraterrestrial planets.</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

Toward accurate extraction of Earth’s free oscillation through a novel deep-learning-based window selection

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Shiyu Zeng,
  • Binbin Liao,
  • Rumeng Guo,
  • Kun Dai,
  • Yijun Zhang,
  • Xiaoming Cui,
  • Jiangcun Zhou,
  • Jianqiao Xu,
  • Xiaodong Chen,
  • Mingqiang Hou,
  • Heping Sun

摘要

The detection and analysis of the Earth’s free oscillations in the form of normal modes contribute to our understanding of the planet’s structure, dynamics, and seismicity. However, false and dismissed recognitions of the focused normal mode signals prevent the detailed exploration of the Earth’s interior due to the complex patterns of the normal mode spectrum. Here, we develop a deep-learning-based neural network, named ModeNet, which is capable of precisely and efficiently selecting the frequency windows to cover the target normal modes on noisy spectra. ModeNet achieves a remarkable precision rate in the discrimination between normal modes and noises of 98.1%, with very few false positives (normal modes misidentified as noises) and false negatives (noises misidentified as normal modes). ModeNet achieves good generalization in processing seismograms with different noise levels, components, and time windows, as well as superconducting-gravimeter (SG) observations. Extracting the frequencies of normal modes within windows selected by ModeNet and further performing a radial (1D) Earth structure inversion, our preferred model shows a better fit to the observed eigenfrequencies than the Preliminary Reference Earth Model (PREM). Therefore, ModeNet would be implemented as a potentially valuable tool for the future 3D structure inversion of the deep Earth and extraterrestrial planets.