Physics-Informed Neural Network Assisted Extraction of Height-Averaged Pulsatile Flow Velocity From Scalar Signal...

Area of Science:

• Fluid Dynamics
• Microfluidics
• Computational Science

Background:

• Accurate flow velocity measurement is crucial for shallow microfluidic velocimetry.
• Conventional scalar imaging velocimetry (SIV) suffers from noise sensitivity.
• Deep neural network-assisted SIV (DNN-SIV) requires extensive labeled data and may lack generalizability.

Purpose of the Study:

• To develop an unsupervised, physics-informed framework for reconstructing pulsatile flow velocity from concentration signals in shallow microfluidic channels.
• To address limitations of conventional SIV and DNN-SIV regarding noise sensitivity and data requirements.
• To enhance the accuracy and robustness of microfluidic flow characterization.

Main Methods:

• Proposed Multiscale Perturbation-Enhanced Physics-Informed Neural Network (MPE-PINN).
• Decomposed scalar transport into steady and pulsatile components.
• Embedded perturbation-based governing equations into the neural network's loss function for unsupervised learning.

Main Results:

• MPE-PINN achieved mean absolute percentage errors (MAPEs) below 1% across a broad operating range (fC ≤ 2 Hz, fQ ≤ 1.6 Hz).
• Demonstrated improved robustness against noisy concentration fields compared to conventional PINN.
• Showcased better preservation of pulsatile velocity features than conventional PINN.

Conclusions:

• MPE-PINN offers an accurate and robust unsupervised approach for extracting pulsatile velocity in shallow microfluidic channels.
• The method has practical potential for microfluidic flow characterization.
• Significant implications for biomedical lab-on-a-chip applications.