Deep Opt: A structure optimization framework based on deep learning, and its application to micromixer optimization

In this work, a new structural optimization framework, Deep Opt, is developed. The framework was built based on deep learning and multi-objective genetic algorithm (NSGA-II) for a simple, novel micromixer optimization process of geometrical structures and operational conditions. First, the process of generating signed distance field (SDF) and computational fluid dynamics (CFD) datasets for different structural micromixers was automated based on Python scripts. Then, the two datasets were reconstructed separately using two encoder–decoder convolutional neural networks (ED-CNN), termed ED-SDF and ED-CFD. The surrogate model for subsequent optimization was created by connecting two neural networks in series. Finally, the optimal micromixer and its flow field distribution were determined by utilizing NSGA-II to perform a multi-objective (minimum pressure drop, maximum mixing index, and minimum mixing energy cost) optimization of the micromixer with two structural variables, the sum of the obstacle radius (Rsum), the ratio of the obstacle radius (Rratio), and a CFD variable Re. Compared with conventional optimization methods, Deep Opt is able to generate high-fidelity flow field for the corresponding structures while optimizing the structural parameters and CFD parameters. In addition, Deep Opt improves the scalability of the optimization process, enabling the customization of optimization targets without the need to reconstruct the dataset, improving the utilization of CFD data. In practice, the framework is not only limited to micromixer optimization but can also be applied to CFD optimization problems with general geometrical configurations, such as the design and optimization of airfoils, stirred tanks, and so on.