Machine learning-enhanced predictive modeling for arbitrary deterministic lateral displacement design and test.

Abstract

The separation of biological particles like cells and macromolecules from liquid samples is vital in clinical medicine, supporting liquid biopsies and diagnostics. Deterministic Lateral Displacement (DLD) is prominent for sorting particles in microfluidics by size. However, the design, fabrication, and testing of DLDs are complex and time-consuming. Researchers typically rely on finite element analysis to predict particle trajectories, which are crucial in evaluating the performance of DLD. Traditional particle trajectory predictions through finite element analysis often inaccurately reflect experimental results due to manufacturing and experimental variabilities. To address this issue, we introduced a machine learning-enhanced approach, combining past experimental data and advanced modeling techniques. Our method, using a dataset of 132 experiments from 40 DLD chips and integrating finite element simulation with a microfluidic-optimized particle simulation algorithm (MOPSA) and a Random Forest model, improves trajectory prediction and critical size determination without physical tests. This enhanced accuracy in simulation across various DLD chips speeds up development. Our model, validated against three DLD chip designs, showed a high correlation between predicted and experimental particle trajectories, streamlining chip development for clinical applications.