Deep learning predicts lift forces in any microfluidic channel geometry
No more retraining for each channel shape—this model generalizes across unseen designs.
Inertial microfluidic devices (IMDs) are a low-cost, high-throughput method for sorting particles and cells, but designing them requires slow numerical simulations to predict particle migration. Recent machine-learning acceleration attempts required training a separate model for each channel cross-section type (e.g., rectangular, triangular), shifting the bottleneck from simulation to training. To break this geometry-specific limitation, Ward-Bond et al. propose a novel deep-learning approach that uses a new set of parameters devoid of geometric descriptors.
The neural network was trained on a diverse set of channel shapes and achieves comparable accuracy to existing models on those geometries. Crucially, it extrapolates effectively to completely unseen cross-sections—such as trapezoidal or custom designs—without retraining. The authors demonstrated that the lift-force model can be plugged into particle tracing simulation software, producing migration patterns consistent with published experimental and numerical results across multiple channel designs. This work paves the way for universal, geometry-free acceleration of microfluidic simulation for lab-on-a-chip applications.
- The model contains no explicit geometric parameters, eliminating the need for shape-specific training.
- It performs comparably to existing models on trained channel shapes but generalizes far better to unseen geometries.
- The pretrained model can be transferred directly into particle tracing simulation software, matching literature results across diverse designs.
Why It Matters
This method could dramatically speed up microfluidic device design for cell sorting and diagnostics.