| A microfluidic mixer is an important application of microfluidic chips,which can quickly achieve efficient mixing between different products in the size of 10-300 micrometers,and is widely used in various fields such as science and engineering.However,the research and development difficulty of microfluidic mixers lies in performance verification,and the experimental verification process takes a long time and the manufacturing cost of mixers is high.Therefore,evaluating the performance of the mixer before manufacturing can help reveal potential defects in the current design.At present,the finite element method is often used to simulate the performance.This method is to establish partial differential equations to model the target system and to achieve the process of modeling the physical field of the target system by discretization of the partial differential equations and solving the discrete equations to obtain numerical solutions.However,this process still has a relatively high research and development threshold and a long research and development cycle.Researchers need to have background knowledge in the field of microfluidics and be proficient in operating professional finite element analysis software.In the first part,artificial intelligence technology is used to explore the performance of microfluidic mixers.Due to the time-consuming finite element simulation,this thesis proposes an artificial neural network model to predict the fluid field of microfluidic mixers from the perspective of accelerated simulation.Specifically,this part proposes a nine-grid network model method to predict the fluid field of the mixer.Using this model to build and train an artificial neural network,the mixing area to be predicted is divided into regular grids,and the fluid field can be rapidly predicted under the condition of only boundary values and geometric modeling.This method can significantly shorten the evaluation time of mixer performance.Taking a typical micromixer as an example,traditional finite element analysis software takes about 10 minutes to simulate,while the nine-grid network model method only takes less than 45 seconds and accelerates by about 15 times.To explore the technical path of computer-aided design of microfluidic mixers based on artificial intelligence algorithms,the second part of the paper constructs a low-cost model based on artificial neural networks and inversion design of two optimization algorithms,achieving an automatic synthesis of structural parameters of microfluidic mixers.These two algorithms are used for the inversion design of the geometric structures of two classic microfluidic mixers,The designed mixer not only maintains high consistency between performance and target performance but also significantly improves efficiency.In addition,this section also explored the efficiency between different algorithms when the accuracy threshold was set to 80%,90%,and 97%.The first mixer used multi-objective genetic algorithms to design a single target in 0.01 seconds,0.0124 seconds,and0.0201 seconds,respectively.The time for designing a single target using the particle swarm optimization algorithm was 0.2175 seconds,12.058 seconds,and 50.577 seconds,respectively.The second mixer uses a multi-objective genetic algorithm to design a single target with a time of 0.007 seconds,0.008 seconds,and 0.019 seconds,respectively,while using particle swarm optimization algorithm to design a single target with a time of 1.850 seconds,3.072 seconds,and 22.270 seconds.In the final chapter of the thesis,a method for automatically synthesizing mixer design parameters based on a deep reinforcement learning algorithm was explored.This section combines the low-cost artificial neural network model trained in the above section to construct a reinforcement learning framework and trains automatic design agents for the two mixers mentioned above.Firstly,the geometric structure of the whole microfluidic mixer is abstracted into a decision model,and then the above-trained artificial neural network model and the design parameter search space of the microfluidic mixer are used to provide an interactive environment for optimal design.Agents calculate reward values,adjust strategies,and finally output the optimal structural design of the mixer according to the current status and actions.The automatic optimization design of two microfluidic mixers trained by a single call of the intelligent agent takes 0.129 seconds and 0.169 seconds,respectively. |
