The realization of proper mixing is crucial in many industrial applications. It can be achieved through highly efficient multi-objective microfluidic and micromechanics devices. A good illustration of mixing is in industrial heat transfer applications or the clinical detection and analysis of various diseases such as cancer and HIV using laboratory on-chip devices that allow for integrating several laboratory functions on a single chip. The usefulness of mixing is also witnessed in developing robust and cost-effective microdetection devices to replace the RT-PCR detection methods for rapid detection of COVID-19 disease and the application of the COVID-19 vaccines. However, due to high design, testing and fabrication related costs, it is often difficult to achieve optimal multi-objective mechanical engineering design. As such, the use of computational tools is highly recommended for mechanical micromixer optimization.
Previous research findings demonstrated that vortex shedding downstream a pillar is an interesting fluid mechanism that we can take advantage for adequate mixing in applied engineering. This can be attributed to its ability to enhance the mixing efficiency and reduce the associated design costs in comparison to the use of complicated zig-zag static channels or active mixing. The effect of different intensities of vortex shedding affect in different manner to the mixing efficiency, which can be enhanced. To this aim, vortex shedding can be controlled and predictable. Random Forest and logistic regression are among the classification and predictive techniques used in machine learning computational approaches. However, none of these techniques have been used before to predict the generation of vortex shedding, despite its practical implication in obtaining optimal mixing.
Herein, Dr. Francisco-Javier Granados-Ortiz and Professor J. Ortega-Casanova from the University of Malaga in Spain proposed a machine learning-aided design optimization (MLADO) approach for designing and optimizing an efficient and inexpensive mechanical micromixer. The design is based on the vortex shedding mechanism obtained when the rectangular pillar structure confined within the microchannel interacts with the two fluids. The prior research objective was to predict the possibility of vortex shedding owing to its significance in enhancing the mixing process. The work is currently published in the journal, Physics of Fluids, and it has been selected as Editor’s Pick.
In their approach, machine learning-based classification algorithms were applied to predict the occurrence of vortex shedding. A trained random forecast classifier was used to predict the geometric configurations that could produce vortex shedding. Also, a multi-objective optimization involving maximizing mixing efficiency and minimizing the energy requirements under some design constraints was investigated. Surrogates, the initial stage of the optimization process, were generated via the Kriging interpolation (Gaussian processes). Furthermore, a non-dominated NSGA-II algorithm was utilized to optimize the design. Finally, the resulting optimal designs were simulated using CFD, and their performances were discussed.
Results showed that the predictive model could infer flow characteristics from limited data. The presence of the confined pillar in the microchannel facilitated the generation of the vortex shedding necessary to enhance the mixing. If the surrogates required extra training data, the random classifier could predict whether the additional points were necessary or not. This helped avoid unnecessary computations, thereby accelerating the process. The resulting micromechanical mixer exhibited superior performance than existing devices, recording a maximum mixing efficiency of approximately 50% even at low Raynolds number, very short microchannel length and very unstable mixing conditions. Additionally, the performance of the simulated optimal candidates agreed with predicted values with negligible differences.
In summary, the study reported the optimal design of a micromechanical device following the suggested MLADO approach. This was an efficient and cost-effective method for designing the device. With a data-driven surrogate creation and the ability to decide whether additional points are necessary or not, it only simulated the configurations leading to vortex shedding. Moreover, the MLADO framework could be automated by including further shape parameterization strategies and by training the classifier with external or lower fidelity data. Based on the results, the resulting micromixer outperformed the existing devices. Furthermore, the NSGA-II algorithm successfully achieved the optimal configuration for the multi-objective problem of enhancing the mixing efficiency and reducing the power consumption. Since the framework is easily extendable to any applied engineering field, the authors said that it would help in the design of more efficient mixing mechanical devices for various practical applications.
Granados-Ortiz, F., & Ortega-Casanova, J. (2021). Machine learning-aided design optimization of a mechanical micromixer. Physics Of Fluids, 33(6), 063604.