Scholarlink

<p>Microfluidics experiments offer high-resolution insights into transport and chemical processes in porous media, yet direct measurement of evolving concentration profiles remains challenging. Numerical simulations can serve as virtual probes but are labor-intensive and computationally expensive. Here, we develop a physics-based machine learning toolbox that transforms such simulations into efficient and scalable virtual probes. Central to our toolbox is the non-intrusive reduced basis method, supported by the U-Net and the Convolutional Autoencoder, which learns mappings from experimental images and physical parameters to concentration profiles. By incorporating physics into its construction, the toolbox delivers accurate predictions with a limited number of training samples. Applied to two microfluidics experiments with different base patterns, the toolbox predicts spatio-temporal concentration profiles, effective diffusivities, and locations with a high probability of precipitation. This paves the way for digital twins that enable real-time analysis and tuning of experiments on the fly.</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-based machine learning toolbox for probing concentration under diffusive regime in microfluidics devices

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月21日

查看全文
  • Ryan Santoso,
  • Yuankai Yang,
  • Mara Lönartz,
  • Jenna Poonoosamy

摘要

Microfluidics experiments offer high-resolution insights into transport and chemical processes in porous media, yet direct measurement of evolving concentration profiles remains challenging. Numerical simulations can serve as virtual probes but are labor-intensive and computationally expensive. Here, we develop a physics-based machine learning toolbox that transforms such simulations into efficient and scalable virtual probes. Central to our toolbox is the non-intrusive reduced basis method, supported by the U-Net and the Convolutional Autoencoder, which learns mappings from experimental images and physical parameters to concentration profiles. By incorporating physics into its construction, the toolbox delivers accurate predictions with a limited number of training samples. Applied to two microfluidics experiments with different base patterns, the toolbox predicts spatio-temporal concentration profiles, effective diffusivities, and locations with a high probability of precipitation. This paves the way for digital twins that enable real-time analysis and tuning of experiments on the fly.