Performance prediction of variable-width microfluidic concentration gradient chips by deep learning

With the development of personalized medicine, customized medications are receiving increasing attention. In order to produce customized medications, it is necessary to prepare medication mixture solutions of specified concentrations. We proposed the design of random variable-width (RVW) microfluidic chips, and predicted their performance through Convolutional Neural Networks. First, a design scheme of RVW microchannel was proposed, and the outlet concentrations and the outlet flow rates were obtained by simulation. Second, the KD-MiniVGGNet model was designed according to the principle of convolutional kernel decomposition. The model was trained with the concentration and flow rate data and predicted the outlet concentration and outlet flow rate for more concentration gradient chips. Finally, an experimental research system was built to verify the accuracy of the prediction results of the KD-MiniVGGNet model. The results showed that the RVW microfluidic concentration gradient chips could widen the range of outlet flow rates by 66.7%. When the query conditions were the same, the RVW concentration gradient chip widened the distribution range of outlet concentration of the three outlets by 9%, 16% and 11%, and the distribution range of outlet velocity of the three outlets by 29%, 28% and 30%, respectively. The accuracy of KD-MiniVGGNet model on the test set of outlet concentrations and flow rates could reach 91.5% and 92.7%, respectively. The average absolute error between the prediction results of KD-MiniVGGNet model and the experimental results was 4.3%. The design method proposed in this study could achieve efficient and accurate design of concentration gradient chips, optimize the performance range of concentration gradient chips, and better offer solution preparation services for pharmaceutical customization. © 2023 Chemical Industry Press. All rights reserved.