Key Laboratory of RF Circuits and Systems, Ministry of Education, and, Zhejiang Provincial Laboratory of Integrated Circuit Design, Hangzhou Dianzi University, China.
Lab Chip. 2021 Jan 21;21(2):296-309. doi: 10.1039/d0lc01158d. Epub 2020 Dec 16.
With the various applications of microfluidics, numerical simulation is highly recommended to verify its performance and reveal potential defects before fabrication. Among all the simulation parameters and simulation tools, the velocity field and concentration profile are the key parts and are generally simulated using finite element analysis (FEA). In our previous work [Wang et al., Lab Chip, 2016, 21, 4212-4219], automated design of microfluidic mixers by pre-generating a random library with the FEA was proposed. However, the duration of the simulation process is time-consuming, while the matching consistency between limited pre-generated designs and user desire is not stable. To address these issues, we inventively transformed the fluid mechanics problem into an image recognition problem and presented a convolutional neural network (CNN)-based technique to predict the fluid behavior of random microfluidic mixers. The pre-generated 10 513 candidate designs in the random library were used in the training process of the CNN, and then 30 757 brand new microfluidic mixer designs were randomly generated, whose performance was predicted by the CNN. Experimental results showed that the CNN method could complete all the predictions in just 10 seconds, which was around 51 600× faster than the previous FEA method. The CNN library was extended to contain 41 270 candidate designs, which has filled up those empty spaces in the fluid velocity versus solute concentration map of the random library, and able to provide more choices and possibilities for user desire. Besides, the quantitative analysis has confirmed the increased compatibility of the CNN library with user desire. In summary, our CNN method not only presents a much faster way of generating a more complete library with candidate mixer designs but also provides a solution for predicting fluid behavior using a machine learning technique.
随着微流控技术的各种应用，数值模拟在制造前验证其性能和揭示潜在缺陷是非常推荐的。在所有的模拟参数和模拟工具中，速度场和浓度分布是关键部分，通常使用有限元分析（FEA）进行模拟。在我们之前的工作[Wang 等人，Lab Chip，2016，21，4212-4219]中，通过使用 FEA 预先生成随机库来自动设计微流混合器。然而，模拟过程的持续时间是耗时的，而有限预生成设计与用户期望之间的匹配一致性并不稳定。为了解决这些问题，我们创造性地将流体力学问题转化为图像识别问题，并提出了一种基于卷积神经网络（CNN）的技术来预测随机微流混合器的流体行为。在 CNN 的训练过程中使用了预生成的 10513 个候选设计的随机库，然后随机生成了 30757 个全新的微流混合器设计，并由 CNN 预测其性能。实验结果表明，CNN 方法仅需 10 秒即可完成所有预测，比之前的 FEA 方法快约 51600 倍。CNN 库扩展到包含 41270 个候选设计，填补了随机库中流体速度与溶质浓度图的空白空间，为用户期望提供了更多的选择和可能性。此外，定量分析证实了 CNN 库与用户期望的兼容性提高。总之，我们的 CNN 方法不仅提供了一种更快的方法来生成更完整的候选混合器设计库，而且还为使用机器学习技术预测流体行为提供了一种解决方案。
