基于深度学习的变宽度浓度梯度芯片性能预测

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

随着个性化医疗的发展,定制化药物受到了越来越多的关注,为了生产定制化药物,需要制备指定浓度的药物混合溶液。本研究首次提出了一种随机变宽度(RVW)结构的微流控浓度梯度芯片,并通过卷积神经网络算法实现芯片的性能预测。首先,设计了一种RVW微流道结构并通过仿真模拟得到出口浓度和出口流速。其次,根据卷积核分解原理设计了KD-MiniVGGNet深度学习模型,使用仿真模拟得到的浓度和流速数据训练模型并预测更多浓度梯度芯片的出口浓度和出口流速。最后,通过实验验证了KD-MiniVGGNet深度学习模型预测结果的准确性。研究结果表明:相较于随机等宽度(REW)浓度梯度芯片,RVW浓度梯度芯片的出口集中流速范围提高了66.7%,三个出口的出口浓度分布范围分别拓宽了9%、16%和11%,三个出口的出口流速分布范围分别拓宽了29%、28%和30%;KD-MiniVGGNet模型在出口浓度和出口流速测试集上的模型准确率分别达到91.5%和92.7%;出口浓度的KD-MiniVGGNet模型预测结果与实验结果之间的平均误差为4.3%。本研究中所提出的设计方法可提高浓度梯度芯片结构的多样性,进一步优化浓度梯度芯片的性能范围,更好地为药物定制提供溶液制备服务。

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.