A framework for adopting machine learning is presented for both analysis and design of microlattices, which can be fabricated using additive manufacturing techniques. Building on graph autoencoders in the deep learning realm, a learning algorithm is designed within an encoder and a decoder (autoencoder), which are responsible for analysis and design of microlattice architectures, respectively. Microlattices are generated by a compact genetic algorithm, and their corresponding mechanical properties are obtained by finite-element analysis. The training dataset consists of 2500 microlattices. The autoencoder is trained in a supervised manner with the graph representation of the generated microlattices. The encoder component learns to infer the mechanical properties of a microlattice as latent variables in the form of force–displacement characteristics, whereas the decoder component is presented with desired mechanical properties as inputs and creates a corresponding microlattice. The decoder is able to generate microlattices from the desired mechanical properties. The decoder-generated microlattices are in good agreement with the original ones within and/or without the training dataset. The ability of the encoder to capture more complex mapping, from microlattice architectures to performance metrics, can be improved by adding more graph convolutional layers to the encoder, i.e., producing deeper networks.

中文翻译：

---

增材制造的深度学习和设计：微晶格架构的框架

提出了一种采用机器学习的框架，用于分析和设计微晶格，可以使用增材制造技术制造。基于深度学习领域的图自动编码器，在编码器和解码器（自动编码器）中设计了一种学习算法，它们分别负责微晶格架构的分析和设计。微晶格由紧凑的遗传算法生成，并通过有限元分析获得其相应的力学性能。训练数据集由 2500 个微晶格组成。自动编码器使用生成的微晶格的图形表示以受监督的方式进行训练。编码器组件学习将微晶格的机械特性推断为力-位移特性形式的潜在变量，而解码器组件以所需的机械特性作为输入呈现并创建相应的微晶格。解码器能够根据所需的机械特性生成微晶格。解码器生成的微晶格与训练数据集内和/或外的原始微晶格非常一致。编码器捕获更复杂映射的能力，从微晶格架构到性能指标，可以通过向编码器添加更多图卷积层来提高，即产生更深的网络。解码器能够根据所需的机械特性生成微晶格。解码器生成的微晶格与训练数据集内和/或外的原始微晶格非常一致。编码器捕获更复杂映射的能力，从微晶格架构到性能指标，可以通过向编码器添加更多图卷积层来提高，即产生更深的网络。解码器能够根据所需的机械特性生成微晶格。解码器生成的微晶格与训练数据集内和/或外的原始微晶格非常一致。编码器捕获更复杂映射的能力，从微晶格架构到性能指标，可以通过向编码器添加更多图卷积层来提高，即产生更深的网络。