Deep learning model predicts particle lift forces across any microfluidic channel shape

A team has published a preprint demonstrating a neural network that predicts inertial lift forces in microfluidic channels without explicit geometry parameters. The model, detailed in arXiv:2605.08109, matches the accuracy of existing machine learning approaches on trained channel shapes while generalizing to unseen cross-sections—rectangular, triangular, or otherwise—without retraining.

Inertial microfluidic devices sort and manipulate particles or cells at high throughput, but simulating particle migration has traditionally required geometry-specific models. Earlier machine learning work sped up these simulations by training separate networks for each channel type, shifting the bottleneck from simulation runtime to training overhead. The new approach sidesteps that by encoding flow and particle data in a way that strips out explicit shape descriptors, letting a single model handle arbitrary geometries.

What stands out

1. 01Single-model coverage. The network performs on par with geometry-specific predecessors when tested on channel shapes it saw during training, but extends to novel cross-sections without additional training runs.
2. 02Particle tracing integration. The authors ported the lift force predictor into particle tracing software and validated migration patterns against published experimental data across multiple channel designs.
3. 03Parameter redesign. Instead of feeding the network explicit width, height, or aspect ratio values, the team constructed a feature set that captures local flow conditions and particle states in a geometry-agnostic way.
4. 04Generalization benchmark. The paper includes test cases on channel shapes absent from the training set, showing prediction errors remain low compared to direct numerical simulation.