Topology design which optimizes the materials or components distributions in a selected domain to achieve certain objectives has been widely applied in engineering field such as transportation vehicles or architectural designs. In structural mechanics, topology optimization (TO) provides systematic approach to optimize the topology of the structure so that desired material properties are fulfilled. However the method requires the iterative calculation of objective function, which involves computationally expensive numerical simulation such as finite element method (FEM) calculation. As the computational cost for the process scales exponentially with the domain or mesh sizes, constant development is performed to improve the efficiency of the method. In recent years, various deep learning...[ Read more ]

Topology design which optimizes the materials or components distributions in a selected domain to achieve certain objectives has been widely applied in engineering field such as transportation vehicles or architectural designs. In structural mechanics, topology optimization (TO) provides systematic approach to optimize the topology of the structure so that desired material properties are fulfilled. However the method requires the iterative calculation of objective function, which involves computationally expensive numerical simulation such as finite element method (FEM) calculation. As the computational cost for the process scales exponentially with the domain or mesh sizes, constant development is performed to improve the efficiency of the method. In recent years, various deep learning-based methods have been developed to accelerate the design process. Although the existing applications of deep learning models could accelerate the design process through different approaches, most of them suffer from the issue of transferability and the requirement of large amount of training data. Since the training data is required to be generated through numerical simulation itself, the overall efficiency of those methods is reduced.

In this thesis, three major ANN-based methods are developed with the ultimate objective of accelerating the traditional topology design process, while also taking into account the transferability of the networks and the training data generation process. In the first method, an inverse design model is built with a combination of a surrogate model based on Convolutional Neural Network (CNN), and a generative model, Deep Convolutional Generative Adversarial Network (DCGAN). The method is developed to solve inverse design problem in a short time, by generating designs that satisfy desired mechanical properties, while also subjected to prescribed geometrical constraint. The design of microstructural materials to achieve specified effective compliance tensor is used as the demonstration for effectiveness of the developed model.

In the second method, a deep learning model known as Mapping Network is developed to reduce the time taken for training data generation of neural network-based surrogate model. Instead of generating all the training data for the surrogate model by performing FEM calculation on the field of interest in the full-scale mesh, a large portion of the data are instead generated in much coarser mesh. The coarse-scale field is then mapped back to the original scale by MapNet. Since the simulation time in coarse-scale mesh is much faster, and the prediction time of Mapping Network is also relatively short, the overall time required during the data generation process can then be greatly reduced. The application of surrogate model in TO process for structural design problem is used to demonstrate the time saving that could be achieved by using the proposed method as compared with the traditional method of training data generation.

Next, using the insights gained from the second method, the idea of Mapping Network in mapping the field of interest from coarse to fine scale is further developed and improved upon. Since the trained network provides a great transferability to different design problems, the idea is integrated to develop a highly scalable surrogate modelling method used for accelerating the most expensive part of TO process, the FEM calculation of objective values. In the developed method, the FEM calculations during each TO step are performed at coarse scale mesh, which are then mapped back to the fine scale using a deep learning model known as MapNet. Combining with the fine scale structure available from each iteration of TO, and fragmentation technique which crops a domain into many smaller subdomains, the trained MapNet has great transferability and can be easily used for different design problems. The developed framework is demonstrated to be easily applied across TO processes for various design problems including structural and thermal problem, while achieving high efficiency and massive time saving.