In this work we created functional microfluidic chips without actually designing them. We accomplished this by first generating a library of thousands of different random microfluidic chip designs, then simulating the behavior of each design on a computer using automated finite element analysis. The simulation results were then saved to a database which a user can query viahttp://random.groverlab.org to find chip designs suitable for a specific task. To demonstrate this functionality, we used our library to select chip designs that generate any three desired concentrations of a solute. We also fabricated and tested 16 chips from the library, confirmed that they function as predicted, and used these chips to perform a cell growth rate assay. This is one of many different applications for randomly-designed microfluidics; in principle, any microfluidic chip that can be simulated could be designed automatically using our method. Using this approach, individuals with no training in microfluidics can obtain custom chip designs for their own unique needs in just a few seconds.

The ability to thoroughly mix two fluids is a fundamental need in microfluidics. While a variety of different microfluidic mixers have been designed by researchers, it remains unknown which (if any) of these mixers are optimal (that is, which designs provide the most thorough mixing with the smallest possible fluidic resistance across the mixer). In this work, we automatically designed and rationally optimized a microfluidic mixer. We accomplished this by first generating a library of thousands of different randomly designed mixers, then using the non-dominated sorting genetic algorithm II (NSGA-II) to optimize the random chips in order to achieve Pareto efficiency. Pareto efficiency is a state of allocation of resources (e.g. driving force) from which it is impossible to reallocate so as to make any one individual criterion better off (e.g. pressure drop) without making at least one individual criterion (e.g. mixing performance) worse off. After 200 generations of evolution, Pareto efficiency was achieved and the Pareto-optimal front was found. We examined designs at the Pareto-optimal front and found several design criteria that enhance the mixing performance of a mixer while minimizing its fluidic resistance; these observations provide new criteria on how to design optimal microfluidic mixers. Additionally, we compared the designs from NSGA-II with some popular microfluidic mixer designs from the literature and found that designs from NSGA-II have lower fluidic resistance with similar mixing performance. As a proof of concept, we fabricated three mixer designs from 200 generations of evolution and one conventional popular mixer design and tested the performance of these four mixers. Using this approach, an optimal design of a passive microfluidic mixer is found and the criteria of designing a passive microfluidic mixer are established.

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.

Microfluidics concentration gradient generators have been widely applied in chemical and biological fields. However, the current gradient generators still have some limitations. In this work, we presented a microfluidic concentration gradient generator with its corresponding manipulation process to generate an arbitrary concentration gradient. Machine-learning techniques and interpolation algorithms were implemented to help researchers instantly analyze the current concentration profile of the gradient generator with different inlet configurations. The proposed method has a 93.71% accuracy rate with a 300× acceleration effect compared to the conventional finite element analysis. In addition, our method shows the potential application of the design automation and computer-aided design of microfluidics by leveraging both artificial neural networks and computer science algorithms.

Automatic optimization of resonant structure is highly desired in the design of high-sensitivity microwave microfluidic sensors. In this paper, the design of resonant structures is abstracted as a decision model, and a deep deterministic policy gradient algorithm based on joint simulation is applied to achieve automatic optimal design through a learned strategy. The agent’s action strategy is decomposed into multiple movement actions against the pixelated structure and finally outputs the optimized structure. Through the optimal structure adjustment strategy, the sensor sensitivity can be dramatically improved. Depending on the design requirement, the liquid consumption volume can be set to be a constant or variable. To evaluate the performance of the proposed strategy, two optimized prototypes are prepared and tested. Compared with the original complementary split-ring resonator-based sensor which obtains a sensitivity of 0.522% for water measurement, the optimized sensors achieve high sensitivities of 0.666% and 0.805%, respectively, implying that the deep deterministic policy gradient-based agent can effectively explore the optimization strategy. This study is a meaningful attempt to develop automatic design procedure for planar microwave microfluidic sensors.

According to the complexity of human physiology and variability among individuals, e.g., genes, environment, lifestyle exposures, etc., personalized medicine aims to synthesize the specific efficacious drug for each individual patient. For synthesizing personalized medicine, customized solutions with specific concentrations are required. Equipped with the advantages in saving costly reagents and rare samples, microfluidic biochips are promising in generating different concentrations for personalized medicine. On the one hand, digital microfluidic biochips require the programming control for driving the movement of the droplets, which suffer from random errors caused by imbalanced droplet splitting. On the other hand, existing flow-based microfluidic biochips can only generate linear concentration gradients, which cause significant waste for synthesizing personalized medicine. To address the above issues, this paper proposes the first artificial neural network (ANN)-based design method for flow-based microfluidic biochips, which accurately generates the customized concentration gradients. According to the required concentration, an initial chip is first selected from the prebuilt database and then fine-tuned by ANN to better match the required concentration. Computational simulation results show that the induced deviations in generated concentrations are generally less than 0.014, which validates the accuracy of the proposed neural network model.