Transfer Learning-Based Microfluidic Design System for Concentration Generation∗

Due to the complexity of human physiology and variability among individuals, e.g., genes, environment, lifestyle exposures, etc., personalized medicine has attracted great interest in the past few years. For synthesizing personalized medicine, it is critical to prepare customized samples with specific concentrations by microfluidic biochips because of the advantages in saving costly reagents and rare samples. The current state-of-the-art of concentration generation for microfluidic biochips is to construct a database by random design methods. However, due to the complex multidimentional parameters such as molecule diameters, inlets, outlets, etc, the whole process is error prone and time consuming. To speedup database construction and reduce the errors in concentration generation, this paper proposes the first transfer learning-based method based on an artificial neural network model (ANN). Given an initial ANN model, transfer learning method can fine-tune weights of ANN to obtain all ANN models needed in the database, which can significantly reduce the amount of required training data. Computational simulation results show that the time for database construction is reduced from several months to 2 days, and the query error is reduced by 83% compared with the existing method.