摘要
Topology optimization (TO) is a powerful computational tool for the design of high-performance structures by optimally distributing material within a prescribed domain. It has enabled significant progress in various engineering applications, including lightweight structural design in aerospace and automotive industries. In these applications, achieving near-optimal performance often requires extremely high-resolution designs with billions of finite elements. However, conventional structural TO methods suffer from prohibitively high computational costs due to repeated high-dimensional simulations required in each optimization iteration. This not only leads to inefficiencies but also demands massive power consumption, which hinders their deployment in real-world design scenarios. This thesis presents a set of efficient structural TO frameworks empowered by artificial neural networks (ANNs), aiming to drastically reduce the computational burden while maintaining design accuracy. Two key categories of TO problems are considered: (1) scenarios that require generating multiple similar designs, and (2) large-scale, high-resolution design tasks. A fundamental distinction between these two lies in data availability-ample training data can be collected for the first type, making it suitable for offline learning, whereas the second type typically lacks sufficient data. To address the first category, a coarse-to-fine structure mapping strategy is introduced, where response calculations are conducted on a coarse grid. A structure mapping neural network (SMapNN) is trained to reconstruct the fine grid design from the coarse grid design at each iteration, thus eliminating the need for expensive fine grid simulations. Fragmentation and padding techniques are further introduced to enhance the generalization capability of the SMapNN. 3D compliance design results demonstrate that, on a high-resolution mesh with over 2 million finite elements, the proposed approach achieves about 23 times speedup while maintaining over 99% compliance accuracy. For large-scale TO where generating extensive training data is challenging, online learning methods present a promising solution. These methods enable adaptive training of surrogate models during the optimization process by leveraging data from previous iterations. However, current online learning approaches employ a coarse-to-fine localized mapping strategy that may compromise surrogate model accuracy due to insufficient global information. Our second framework aims to overcome the undetermined mapping issue of existing online learning TO methods. A mapping between local patches and sensitivities is established by employing the Fourier Neural Operator (FNO) as a surrogate model. By embedding global context into local patches, the mapping uniqueness is significantly improved. The training and inference processes are performed entirely on a single-resolution mesh, avoiding costly coarse-grid simulations of the existing method. The results of numerical experiments demonstrate the effectiveness of the proposed framework across various compliance design problems-for instance, it achieves a 4.3-fold speedup in a 3D cantilever design case with about 0.5 million resolutions. Lastly, this approach is further extended into the isogeometric analysis framework, enabling accurate geometric representation and improved stress evaluation. We then demonstrate the framework’s effectiveness through several 2D stress-constrained design cases, where the method achieves a 1.8 times speedup on a problem with about 20 thousands control points, and greater acceleration is expected for large-scale 3D designs. To address the relatively high training and inference cost of the FNO, a decomposed FNO (D-FNO) is finally proposed. By decomposing high-dimensional latent features into a series of rank-1 tensor products, the traditional three-dimensional (3D) fast Fourier transform (FFT) is replaced with a series of one-dimensional FFTs, reducing the computational complexity from 𝑂(𝑁3log𝑁) to 𝑂(𝑃𝑁log𝑁) in 3D problems, where 𝑃 is the decomposition term. The D- FNO is validated across various PDE scenarios. In the most challenging 3D elastic wave propagation experiment, the D-FNO reduces the training time by up to 7.6 times and inference time by 7.0 times, while maintaining a slightly better prediction accuracy. Finally, its effectiveness is further demonstrated in 3D online learning TO tasks. It achieves a 4.6-fold speedup in a cantilever design case with about 0.5 million resolutions, showing 8.2% efficiency improvement compared to the standard FNO.