三周期极小曲面（Triply Periodic Minimal Surface，TPMS）因其优异的性能在工程领域引起了广泛关注。但由于设计过程复杂，应用仍然面临一定限制。深度学习的快速发展，为超材料结构设计带来了新的可能，本文基于深度学习方法，提出了一种TPMS结构的性能预测与逆设计方法，旨在实现可调控的力学性能、结构可操控性，并探讨其在汽车领域的应用。

首先介绍了TPMS和深度学习方法在微结构逆向设计方面的应用与发展过程，采用近似函数方法构建了TPMS元胞。通过有限元仿真计算了其力学性能，并利用增材制造技术制备了TPMS样件进行压缩实验，验证了仿真的有效性。随后通过调整TPMS元胞的结构参数，构建了大量的元胞结构用于仿真计算。将这些元胞转化为体素或点云，并将其与计算得到的等效弹性模量一一对应。通过数据增强等方法，分别构建了用于性能预测和逆向设计网络训练的数据集。

然后针对传统正向设计方法过于依赖于研究人员经验、有限元模拟和实验验证，耗时且方向不确定等问题。提出了使用深度学习方法，具体包括3D Convolutional Neural Networks（3DCNN）和Warping Generative Adversarial Networks（WarpingGAN），用于实现TPMS结构的等效弹性模量预测和逆向设计。训练好的3DCNN可以快速且准确地预测TPMS元胞结构的性能，本文将预测结果扩充到点云数据集中用于逆设计网络的训练。另外，本文的逆向设计方法的输出结果可以通过简单的点云重建，从网络输出中直接获得所需的结构，避免了复杂的参数化建模过程。

最后将逆向设计生成的TPMS元胞应用在了汽车动力电池的防护设计上，探讨了TPMS结构在侧边梁碰撞防护中的作用。结果表明，逆向设计出的在2000MPa左右的元胞结构就可以实现足够的吸能，在这三种元胞结构中Primitive表现最佳。若是选用更大的等效弹性模量的元胞则会在一定程度上让最大加速度、最大峰值力、单位质量吸能等指标的值变差，但是也为更高速度的碰撞预留了空间。将这些逆向设计的元胞填充到侧边梁内，可以有效吸收侧碰时的冲击，提高车辆的安全性。

Triply Periodic Minimal Surfaces (TPMS) have attracted widespread attention in the engineering field due to their excellent performance. However, their application is still limited due to the complexity of the design process. The rapid development of deep learning has opened up new possibilities for the design of metamaterial structures. This paper presents a deep learning-based method for performance prediction and inverse design of TPMS structures, aiming to achieve tunable mechanical properties and structural controllability, and explores its application in the automotive industry.

Firstly, the paper introduces the application and development of TPMS and deep learning methods in microstructure inverse design. An implicit function method is used to construct TPMS cells. The mechanical properties of these cells are computed through finite element simulations, and TPMS samples are fabricated using additive manufacturing technology to conduct compression experiments, validating the effectiveness of the simulations. Later, by adjusting the structural parameters of the TPMS cells, a large number of cell structures are constructed for simulation calculations. These cells are converted into voxels or point clouds, which are then matched with the computed equivalent elastic modulus. Using data augmentation techniques, a dataset is constructed for training the performance prediction and inverse design networks.

Then, addressing the issues of traditional forward design methods, which rely heavily on researchers' experience, finite element simulations, and experimental validation, the paper highlights the challenges of these approaches. These traditional methods often result in time-consuming processes and uncertain outcomes. To overcome these limitations, the paper proposes a deep learning approach. Specifically, it utilizes 3DCNN and WarpingGAN for predicting the equivalent elastic modulus and performing inverse design of TPMS cell structures. The trained 3DCNN can quickly and accurately predict the performance of TPMS structures. The predicted results are expanded into a point cloud dataset for training the inverse design network. Furthermore, the output of the inverse design method can be directly retrieved from the network's output through simple point cloud reconstruction, eliminating the need for complex parametric modeling.

Finally, the generated TPMS cells are applied to the protective design of automotive power batteries, exploring the role of TPMS structures in side impact protection. The results show that TPMS cells with an equivalent elastic modulus around 2000 MPa can provide sufficient energy absorption, with the Primitive cell structure performing the best among the three types tested. If cells with a higher equivalent elastic modulus are selected, the values of indicators such as Maximum Acceleration, Peak Crushing Force, and Special Energy Absorption may decrease to some extent, but they also allow for higher-speed collision scenarios. Filling these reverse-designed cells into the side beam can effectively absorb the impact during a side collision, enhancing the vehicle's safety.