$μ$-FlowNet: A Deep Learning Approach for Mapping Flow Fields in Irregular Microchannels Using an Attention-based U-Net Encoder-Decoder Architecture

Abstract: In the complex domain of microfluidics systems, analysing fluid flow patterns through random-shaped circular microchannels is significantly challenging task. Conventional approach of solving such problems using computational fluid dynamics often incapable due to their intensive computational requirements and high simulation times. In this study, addressing these limitations, we introduce $μ$-FlowNet, a deep learning framework based on the adaptable U-Net autoencoders. This model provides a data-driven approach that enhances the prediction and mapping of random-shaped circular microchannels and their corresponding fluid flow patterns. The datasets required for the training of the model is generated by performing extensive simulations using conventional approach of computational fluid dynamics methods. The datasets are then pre-processed and accessed the required spatial and temporal features that are essential for the training. We have trained three different models based on U-Net framework namely, standard U-Net, T-Net, and U-Net with attention mechanism to compare the prediction accuracy and loss. The accuracy of the $μ$-FlowNet is compared using metrics of dice score and intersection over union and it shows that U-Net with attention mechanism shows the highest dice score and IoU of 0.9317 and 0.8731, respectively and shows the highest structural similarity as compared to standard U-Net and T-Net. This show that U-Net with attention mechanism serves best model to map the fluid flow pattern with random datasets on testing.

• The paper introduces μ-FlowNet, an attention-based U-Net that accurately maps complex microchannel geometries to steady-state flow fields.
• It integrates CFD simulation data with Weierstrass-Mandelbrot-generated geometries, achieving Dice scores above 0.92 in flow prediction.
• The model drastically reduces computational time (4.6 ms vs 300,000 ms per sample), enabling real-time design and inverse optimization in microfluidics.

Deep Learning-Based Flow Field Mapping in Irregular Microchannels: The $μ$-FlowNet Framework

Introduction

Accurate prediction of fluid flow in irregular, rough-walled microchannels is essential for microfluidics, biomedical device engineering, and the optimization of lab-on-a-chip systems. Classical CFD approaches, while precise, incur substantial computational cost and are not feasible for rapid prototyping or high-throughput inverse design. The paper "$μ$-FlowNet: A Deep Learning Approach for Mapping Flow Fields in Irregular Microchannels Using an Attention-based U-Net Encoder-Decoder Architecture" (2604.17575) systematically addresses these challenges by leveraging data-driven surrogate modeling. Specifically, it proposes an attention-augmented U-Net framework (μ-FlowNet) trained to map complex random-shaped channel geometries, modeled using Weierstrass-Mandelbrot fractal theory, to corresponding steady-state velocity fields.

Methodology

Dataset Generation and Physical Modeling

The dataset construction begins with generation of 1334 random microchannel geometries using a parametrized Weierstrass-Mandelbrot function, capturing multi-scale roughness characteristics representative of natural and fabricated microfluidic environments. Corresponding steady 2D incompressible laminar flows are solved numerically via finite element-based CFD (COMSOL). Boundary conditions include fixed mean inlet velocity and no-slip at channel walls, with the simulations yielding pixel-wise velocity vectors that serve as training targets.

Data Representation

Channel geometries are encoded as binary images (fluid/solid regions) on a $128 \times 256$ Cartesian grid for input to the networks, with output target maps for both $u$/$v$ velocity components and the velocity magnitude. This formulation casts flow field prediction as an image-to-image regression problem.

Deep Learning Architectures

Three architectures are benchmarked:

• Standard U-Net: An encoder-decoder CNN with skip connections for spatial localization.
• T-Net: A fully convolutional model employing sequential upsampling blocks and asymmetric kernels, originally designed for vessel segmentation tasks.
• Attention U-Net (μ-FlowNet): The canonical U-Net enhanced with attention modules—specifically, Channel Attention Modules (CAM) and Spatial Attention Modules (SAM)—inserted at skip connections and the network bottleneck. These modules prioritize salient spatial and channel features, improving the ability to capture fine-grained boundary-driven flow phenomena.

All models are trained using L1 loss augmented with a physical loss function to enforce predictions that are physically plausible, and optimized using Adam.

Evaluation Protocol

Model accuracy is assessed using multiple metrics:

• Mean Relative Error (MRE)
• Dice Similarity Coefficient
• Intersection over Union (IoU)

Comparisons are made over train/validation splits, and prediction runtime versus CFD simulation time is also reported.

Results

Numerical Performance

The attention-based μ-FlowNet exhibits the strongest predictive performance:

• For velocity magnitude, Dice score 0.9264 and IoU 0.8628.
• X-velocity (u) component, Dice score 0.9269; Y-velocity (v) component, Dice score 0.9317.
• Mean relative errors for all metrics are consistently lower than both T-Net and vanilla U-Net.

Most notably, μ-FlowNet achieves substantial acceleration, being around 64,946 times faster in inference than full CFD simulation (4.6 ms vs 300,000 ms per sample), with T-Net and U-Net also achieving multiple orders of magnitude improvement.

Qualitative Flow Field Fidelity

Across a variety of challenging geometries, μ-FlowNet outputs exhibit high-fidelity prediction, particularly in regions of steep velocity gradients and near boundary irregularities where classical models typically struggle. The attention gates improve spatial focus, minimizing errors in the vicinity of sharp protrusions or highly non-uniform channel contours. Error heatmaps show that most discrepancies are highly localized and attenuated compared to non-attention models.

Model Efficiency and Trade-offs

While the Attention U-Net has more parameters (37M vs 9.3M for U-Net and 4.3M for T-Net), the increase in runtime remains negligible relative to CFD. The model stably converges within 100 epochs for all tasks. T-Net offers competitive accuracy, shorter runtime, and leaner architecture, but cannot match the accuracy or boundary sensitivity achieved by the attention-augmented version.

Implications and Future Directions

The integration of attention mechanisms into U-Net-derived architectures represents a significant advancement for surrogate modeling of fluid dynamics in complex geometries. The demonstrated speed-up democratizes CFD-based exploration by enabling real-time design iteration, rapid inverse design, and integration of microfluidic simulation in closed-loop optimization pipelines or digital twins.

The practical implications are extensive for:

• Rapid prototyping of microfluidic chips with intricate channel topologies.
• Data-driven inverse optimization, directly linking geometry parameters to performance metrics for applications such as enhanced mixing or targeted flow control.
• Coupling with downstream physiochemical modules (e.g., mass transport, cell deformation models) for biologically relevant flows.

Theoretical Impact

The use of fractal (Weierstrass-Mandelbrot) boundary specification aligns with physical roughness models, moving surrogate modeling away from simplistic synthetic benchmarks and toward realistic, experimentally observed domains. The careful construction of attention modules, focusing on spatial and channel-wise saliency, points toward further research into task-specific architectural innovations for PDE surrogate learning.

Extensions

The paper outlines several compelling directions:

• Extension to 3D turbulent and multiphase flow problems by incorporating physics-informed loss terms or hybridizing with PINNs.
• Domain generalization via federated learning, aggregating data across heterogeneous microfluidic platforms.
• Use as a differentiable surrogate for inverse geometric optimization (e.g., microchannel design for optimized shear or mixing).
• Combination with adversarial learning to generate novel microfluidic architectures beyond classical engineering intuition.

Conclusion

μ-FlowNet establishes an efficient, accurate, and scalable framework for predicting flow fields in random rough-walled microchannels, significantly outperforming both traditional CFD and plain deep image-to-image architectures. Its attention-based design ensures precise boundary-dependent mapping, which is critical given the sensitivity of microfluidic phenomena to geometric perturbations. The rapid inference speed, high accuracy, and extensibility position μ-FlowNet and similar architectures as fundamental building blocks for intelligent microfluidics and data-driven design applications. Continual advances in geometric deep learning and physics-constrained surrogate modeling are poised to further accelerate innovation in this domain.

Reference: "$μ$-FlowNet: A Deep Learning Approach for Mapping Flow Fields in Irregular Microchannels Using an Attention-based U-Net Encoder-Decoder Architecture" (2604.17575)