Portal de Periódicos da CAPES

Artigo Acesso aberto Produção Nacional Revisado por pares

BIBTEX RIS

Computational fluid dynamics and machine learning as tools for optimization of micromixers geometry

2022; Elsevier BV; Volume: 194; Linguagem: Inglês

10.1016/j.ijheatmasstransfer.2022.123110

1879-2189

Daniela de Oliveira Maionchi, Luca Ainstein, Fabio Pereira dos Santos, Maurı́cio Bezerra de Souza Júnior,

Microfluidic and Bio-sensing Technologies

Microfluidic devices have become a new trend in different fields and have attracted attention due to their compact size and capability to deal with a small amount of fluid. Micromixing is an efficient way to mix miscible fluids at this microfluidic level. This work explores a new approach for optimization in microfluidics, using CFD (Computational Fluid Dynamics) and ML (Machine Learning) techniques. The objective of this combination is to enable global optimization with lower computational costs. A Y-type micromixer inspires the initial geometry with cylindrical grooves on the surface of the main channel and obstructions inside it. Simulations for circular obstructions were carried out using the OpenFOAM software to observe the influence of these obstacles. The effects of obstruction diameter and its offset on the percentage of mixing, pressure drop and energy cost were investigated. Numerical experiments were analyzed using machine learning. A neural network was used to train the dataset composed of obstruction diameter and its offset as inputs and percentage of mixing and pressure drop as outputs. The genetic algorithm was used to find the geometry that offers the maximum value of the percentage of mixing and the minimum pressure drop value. The optimal value obtained for the obstruction diameter was 131 mm and for its offset 10 mm, which corresponds to obstructions of medium size close to the channel wall. It is worth mentioning that each simulation takes around 4h, and the total time to guarantee the global optimization would be about 25.000 days. With this methodology, the dataset production, training and optimization takes 40 days. This procedure is a tremendous advantage for microfluidic optimization.