Machine Learning-Driven Prediction of DLD Chip Throughput

Abstract

Abstract The microfluidic chip technology, capable of manipulating fluids at the micrometer-scale, is increasingly being applied in the fields of cell biology, molecular biology, chemistry, and life sciences. The densely integrated microfluidic chip devices enable high-throughput parallel experiments and integration of various operational units. However, the development of densely integrated microfluidic chips also comes with high demands on driving equipment. Due to manufacturing processes and inherent design limitations, the driving capability of the equipment is restricted. To address potential challenges faced by microfluidic chips in the development towards integrated biological microsystems and to maximize their high-throughput performance, improvements are required not only in selecting appropriate driving equipment but also in design aspects. This study focuses on the DLD chip and delves into the complexity of microfluidic chip design. By combining Bézier curves to characterize arbitrarily shaped micropillars and conducting finite element analysis to compute the pressure field of DLD chips, we explore methods utilizing random forest, XGBoost, LightGBM, and ANN machine learning algorithms to predict the impedance of DLD chips. Our objective is to guide engineers in designing chips with smaller impedance (lower pressure drop) and larger throughput more quickly and efficiently during the design phase. Ultimately, through evaluating the predictive capabilities of the four models on new data, we select the ANN algorithm model to predict the pressure drop under different designs of DLD chips. This offers possibilities for enhancing the efficiency and integration of microfluidic technology in biomedical applications.