Design Of Random Variable-Width Microfluidic Concentration Gradient Chip Based On Deep Learning

With the advantage of making pharmaceutical solutions with different concentrations efficiently and accurately,concentration gradient chips have outstanding performances in many fields such as personalized medicine,drug screening,and antibiotic susceptibility testing.Currently,concentration gradient chips suffer from a narrow range of outlet concentration and flow rate distributions and lengthy design cycles.Therefore,there is a need to develop a concentration gradient chip structure with a wider range of outlet concentration and flow rate distributions and to achieve an efficient and accurate design of the concentration gradient chips.The design method of Random variable-width(RVW)concentration gradient chips based on deep learning was proposed for the first time,and the research focused on three aspects,such as mechanism analysis,design method research,and experimental research.The main contents and conclusions of this study are as follows.(1)The mass transfer mechanism of the concentration gradient chip and the architecture of convolutional neural network(CNN)algorithms were analyzed insightfully.The basic characteristics and fundamental equations of microfluidics in the flow process and the mechanism of concentration gradient generation were investigated,and it was found that the outlet fluid behavior of concentration gradient chips could be modified by the combination of microchannels of different widths.The basic features,basic structure,and solution process of the CNN algorithm were analyzed,and it was discovered that the geometric structures of the concentration gradient chips could be represented by a geometric matrix,which could effectively improve the training speed and reduce the FLOPs of the CNN model.(2)The novel structure of the RVW concentration gradient chip was proposed for the first time through the combination of microchannels with different widths.The outlet fluid behavior of various RVW concentration gradient chip structures was investigated by numerical simulation methods.The results showed that compared with the Random equal-width(REW)concentration gradient chips,the concentration distribution range of the three outlets of the RVW concentration gradient chips was widened by 9%,16%,and 11%,and the flow rate distribution range was widened by 29%,28%,and 30%,respectively.The optimal outlet concentration and outlet flow rate distribution of RVW concentration gradient chips were obtained when the number of microchannel nodes was 5×5,the total number of possible microchannels was 40,the probability of microchannel occurrence was 70%,and the number of existing microchannels was 28.(3)The CNN models were trained by the outlet flow rate and outlet concentration datasets obtained from COMSOL simulations to obtain a design method for the RVW concentration gradient chip.The Velocity NET model and Concentration NET model were trained with the data sets obtained by numerical simulation,and the results showed that the prediction accuracy of the models were 92.68% and 91.51%,respectively,achieving accurate prediction of the outlet fluid behavior.Based on transfer learning,two models,c Transfernet and v Transfernet were established to expand the concentration gradient chip database under different inlet conditions.An RVW concentration gradient chip retrieval and design method based on a database was provided,which realized the minute-level design of concentration gradient chip.(4)The accuracy of the Concentration NET model was verified experimentally on a 3Dprinted RVW concentration gradient chip.The average absolute errors between the predicted and experimental values of the outlet concentration were 3.50%,5.17%,and 4.23% for the experiments conducted at different inlet flow rates,different inlet concentrations,and different inlet materials,respectively.The experimental results proved that the method in this study could achieve the automatic design of RVW concentration gradient chips efficiently and accurately.The structures of RVW concentration gradient chips were proposed for the first time,and the design of RVW concentration gradient chips was realized efficiently and accurately by the Velocity NET model and Concentration NET model.Through the mechanistic analysis,the mass transfer mechanism of concentration gradient chips and the structure of the CNN algorithm were analyzed.Through the research on the design method,the key structural parameters of the RVW concentration gradient chips were specified,and a design method based on the CNN model was provided.Through the experimental research,the accuracy of the RVW concentration gradient chip design method proposed in this study was demonstrated.The results indicate that the RVW concentration gradient chips designed based on the CNN model have the advantages of a wider range of outlet concentration and outlet flow rate,high design efficiency and accuracy,and can be applied to the automatic design of other microfluidic chips.