Data-driven design of deterministic lateral displacement microfluidic devices for CTC separation

Circulating tumor cells (CTCs) are established biomarkers for cancer diagnosis and therapeutic monitoring, yet their extreme rarity in peripheral blood and size overlaps with leukocytes present substantial technical barriers to clinical deployment. Deterministic lateral displacement (DLD) is a promising microfluidic technique for label-free, continuous CTC separation. However, due to the extreme heterogeneity of cancer cells, rational device design requires computationally expensive iterative simulation across multi-dimensional geometric parameter spaces. Here we present a data-driven design framework that integrates validated computational fluid dynamics (CFD) simulations with supervised machine learning to enable rapid prediction of particle separation behavior and critical diameter in DLD arrays. A dataset of approximately 8.5 million trajectory data points was generated across 1160 unique combinations of array period number (N = 3-48) and particle diameter (1-20 μm) using COMSOL-based CFD. Four regression algorithms, Gradient Boosting, Random Forest, k-Nearest Neighbors, and Multi-Layer Perceptron, were trained and evaluated on trajectory prediction and critical diameter estimation. Random Forest achieved the highest accuracy, with a test-set R² of 0.994 and a mean absolute error of 0.31 μm in critical diameter prediction. A clinical case study targeting colorectal CTC separation from blood demonstrated that the ML-guided design workflow converged on an optimal device configuration (N = 15, D_c = 10-12 μm) in under 2.3 s, representing a speedup exceeding four orders of magnitude compared to equivalent CFD iteration. The framework enables direct translation of CTC size distributions into DLD device configurations and provides a scalable foundation for reproducible microfluidic process development in biomedical applications.