Microfluidic advancements heavily depend on precisely engineered 3D flow patterns for microscale sample manipulation. While numerical methods offer insights, accurately simulating microflows remains challenging due to complex driving forces, geometries, and fluid-surface interactions. Experimental validation of flow topologies is crucial, demanding high-resolution 3D flow mapping techniques. Micro-particle image velocimetry (µPIV), considered the gold standard, cross-correlates particle patterns, inherently limiting its 3D resolution to several micrometers. Improving µPIV resolution involves costly and complex multi-camera setups or specialized confocal illumination, hindering widespread adoption. Alternatively, single-particle tracking offers sub-pixel 2D resolution, expandable to 3D by comparing particle images to a defocused reference library. Defocusing micro-particle tracking velocimetry (µPTV) employs this approach, often using fluorescence microscopy with specialized optical components (3-pinhole aperture or cylindrical lens). However, these methods suffer from low signal, further reduced by the optical components and low numerical aperture (NA) objectives, leading to limited temporal resolution and complex setups. This research introduces a high-resolution µPTV method using a simple brightfield microscopy setup, open-source software (Fiji and TrackMate), and a modular workflow allowing for cross-correlation or deep learning-based classification for Z-position determination. The study validates the method through three microfluidic examples: channel step expansion, displacement structures (single-phase flow), and droplet microfluidics (two-phase flow), showcasing its potential for efficient design of novel microfluidic structures.

The existing literature highlights the limitations of current 3D microflow mapping techniques. µPIV, while a gold standard, suffers from limited Z-resolution due to its 2D nature and the need for computationally expensive calculations to infer flow between slices. High-resolution µPIV methods require expensive and complex setups involving multiple cameras or specialized confocal systems. Single-particle tracking offers an alternative, with the possibility of extending 2D resolution to 3D using defocusing techniques. However, most µPTV implementations rely on fluorescence microscopy with additional optical elements, resulting in low light signals, limited temporal resolution, and complex setups. The authors review these limitations and position their approach as a simpler, more widely accessible alternative utilizing brightfield microscopy and open-source software.

The proposed method uses brightfield microscopy and open-source algorithms (Fiji and TrackMate) for high-resolution 3D microflow mapping. It employs a two-step strategy: 1) 2D particle tracking using TrackMate in ImageJ to obtain lateral (XY) trajectories; and 2) Z-position determination by classifying defocused particle images against a reference library. The optimal defocusing pattern was achieved by adjusting the correction ring of the objective lens. 3 µm polystyrene microbeads were used as seed particles. The particle density was carefully chosen to balance data density and manual correction needs. The Z-position classification was performed using both a traditional cross-correlation method and a deep learning model, with their performance compared. Accuracy was assessed using synthetic images, measuring root mean square error (RMSE) between predicted and actual Z-positions. Precision was estimated using experimental data from a Poiseuille flow in a straight channel, determining the median RMSE along pathlines. The impact of re-labeling training images on deep learning model accuracy was investigated. The method was validated through three microfluidic experiments: a channel step expansion, displacement structures, and droplet microfluidics. Continuity error analysis was conducted to assess the physical validity of the velocity fields. The computational costs of both the cross-correlation and deep learning approaches were compared.

The study demonstrates a high-accuracy and high-precision 3D microflow mapping technique using brightfield microscopy and open-source software. The deep learning model showed improved accuracy (0.41 µm RMSE) compared to cross-correlation (0.63 µm RMSE) for Z-position prediction. Re-labeling training images was crucial for optimal deep learning model performance. The method successfully mapped the flow field in a channel step expansion, revealing details such as the parabolic velocity profile upstream and out-of-plane flow downstream. Continuity error analysis confirmed the physical validity of the velocity field. The application of the method to displacement structures elucidated their working principles by revealing how they efficiently shift particles across streamlines. In droplet microfluidics, the method revealed novel recirculation structures and folding patterns in the internal flow of droplets. The deep learning model showed a significant speed advantage (two orders of magnitude faster) compared to the cross-correlation method, particularly beneficial for large datasets or iterative design optimization. The method's precision in Z-position was comparable to the precision in the XY plane (around 0.15µm).

The developed method successfully addresses the limitations of existing 3D microflow mapping techniques by providing a simple, accessible, and high-resolution approach. The use of brightfield microscopy eliminates the need for fluorescence and specialized optics, improving temporal resolution and reducing cost. The modular workflow and use of open-source software further enhance accessibility. The validation experiments showcase the method's ability to reveal intricate flow details in various microfluidic scenarios. The superior accuracy and computational efficiency of the deep learning model make it well-suited for complex experiments and iterative design optimization. The findings have implications for advancing microfluidic device design and optimization, leading to more efficient and controlled manipulation of microscale samples.

This study presents a widely accessible and efficient method for 3D microflow mapping at high spatial and temporal resolutions using brightfield microscopy, open-source software, and a choice of cross-correlation or deep learning for Z-position classification. The method demonstrated high accuracy and precision, revealing detailed flow patterns in diverse microfluidic applications. The significant speed advantage of the deep learning approach offers practical benefits for large-scale analyses and iterative design processes. Future work could explore the application of this method to more complex microfluidic systems and further refine the deep learning model for even greater accuracy and efficiency.

While the method demonstrates high accuracy and precision, some limitations exist. The near-wall flow is difficult to capture due to the limitations of particle-based methods. The accuracy of the deep learning model is sensitive to labeling errors in the training dataset, emphasizing the need for careful data curation and potential bias correction. Furthermore, the method relies on the assumption of incompressible flow for continuity error analysis. The computational cost analysis is specific to the hardware and software used and may differ with other setups.