Performance prediction of random variable-width microfluidic chips by convolutional neural networks

With the development of personalized healthcare, tailor-made medications are receiving increasing attention. Solutions of specific concentrations or flow rates need to be acquired before medication can be manufactured. To efficiently and accurately generate solutions with specific concentrations or flow rates, we proposed the design of random variable-width (RVW) microfluidic chips, which perform significantly outperform random equal-width (REW) microfluidic chips, and predict their performance through Convolutional Neural Networks (CNN). First, we proposed the design of RVW microfluidic chips to extend the range of concentrations and flow rates. Second, the KD-MiniVGGNet model was designed, which effectively improved the accuracy of predicting the outlet concentrations and flow rates of the RVW microfluidic chips. Finally, a database of 51 032 RVW microfluidic chips was built by the KD-MiniVGGNet, which provided a sufficient number of candidate designs. The results showed that the RVW microfluidic chip could provide broader and better candidate designs, and the prediction accuracy of the outlet fluid behavior could be increased to 93%.