A CFD-driven supervised learning framework for rapid DLD microfluidic chip optimization in CTC separation

Tanbir Sarowar

doi:10.1007/s10544-026-00813-1

A CFD-driven supervised learning framework for rapid DLD microfluidic chip optimization in CTC separation

2026-04-23

0

PRE

AI

DOI：10.1007/s10544-026-00813-1

Original

Original request for help

Save

Abstract

En 中文

Circulating tumor cells (CTCs) provide critical diagnostic information for cancer detection and monitoring, yet their isolation from whole blood remains technically challenging due to their extreme rarity. Deterministic lateral displacement (DLD) microfluidics offers label-free, size-based separation but requires precise geometric optimization, a process traditionally requiring complex and computationally expensive iterative simulations. Here we present a computational supervised machine learning framework, trained on validated CFD simulation data, that accelerates DLD device design by over three orders of magnitude compared to iterative CFD-based optimization. We generated 8.9 million particle trajectory data points through validated computational fluid dynamics simulations spanning 896 device configurations, systematically varying period number (N = 3–48) and particle diameter (1–14 μm). Four supervised regression algorithms, Gradient Boosting, k-Nearest Neighbors, Random Forest, and Multi-Layer Perceptron were trained to predict particle trajectories as functions of design parameters. Random Forest achieved the highest predictive accuracy (R² = 0.958) with an inference time of 89 ms per design candidate, enabling real-time interactive design exploration. The models, through prediction of the particle trajectories, successfully captured the deterministic zigzag-to-bumped mode transition and accurately identified the critical diameter range. This data-driven approach significantly reduces the reliance on repeated CFD simulations during design optimization, compressing the design exploration cycle from weeks to seconds while maintaining accuracy consistent with the underlying validated simulations. The presented framework provides a generalizable computational methodology that integrates physics-based simulations with supervised learning to accelerate microfluidic device design, offering a potential foundation for future clinical diagnostic applications.

Keywords:

Circulating tumor cells

Deterministic lateral displacement

Microfluidics

Supervised learning

Computational fluid dynamics

Cancer cell separation

A CFD-driven supervised learning framework for rapid DLD microfluidic chip optimization in CTC separation

Abstract

Journal

Organization

Cited Papers