Scholarlink

<p>In inverse imaging and scattering problems, it is critical to avoid potentially solving problems with ambiguity, where the solution space may exceed what is represented in observation data that could lead to non-uniqueness. Thus, inspecting such an issue before collecting data from real-world experiments would be highly desirable. Here, starting with an elastic wave imaging problem aimed at extracting material and geometry parameters, we handle this issue by adopting an unsupervised machine learning technique, variational autoencoder, to compress the observation data into a latent space whose dimensionality is leveraged to assess whether it is unique or not in the inverse process, based on readily accessible data from simulation configured to match experimental settings. After confirming the uniqueness, this latent representation, through independent component analysis, is then converted into independent components each corresponding to one physical parameter to be recovered. The practical application is demonstrated by applying the trained machine learning model to the experimental data measured from 3D-printed samples, with high accuracy. Our method offers broad applications ranging from evaluating the adequacy of data in inverse problems, extracting independent representations from complex data, to building understanding from observations.</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

Inverse imaging with elastic waves driven by unsupervised machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月10日

查看全文
  • Liyou Luo,
  • Yaxi Shen,
  • Jiawei Xi,
  • Yabin Jin,
  • Daniel Torrent,
  • Jensen Li

摘要

In inverse imaging and scattering problems, it is critical to avoid potentially solving problems with ambiguity, where the solution space may exceed what is represented in observation data that could lead to non-uniqueness. Thus, inspecting such an issue before collecting data from real-world experiments would be highly desirable. Here, starting with an elastic wave imaging problem aimed at extracting material and geometry parameters, we handle this issue by adopting an unsupervised machine learning technique, variational autoencoder, to compress the observation data into a latent space whose dimensionality is leveraged to assess whether it is unique or not in the inverse process, based on readily accessible data from simulation configured to match experimental settings. After confirming the uniqueness, this latent representation, through independent component analysis, is then converted into independent components each corresponding to one physical parameter to be recovered. The practical application is demonstrated by applying the trained machine learning model to the experimental data measured from 3D-printed samples, with high accuracy. Our method offers broad applications ranging from evaluating the adequacy of data in inverse problems, extracting independent representations from complex data, to building understanding from observations.