In recent years, metal cellular structures have drawn attentions in various industrial sectors due to their design freedoms and abilities to achieve multi-functional mechanical properties. However, metal cellular structures are difficult to fabricate due to their complex geometries, even with modern additive manufacturing technologies such as the direct metal laser sintering (DMLS) process. Assessing the manufacturability of metal cellular structures via a DMLS process is a challenging task as the geometric features of the structures are complex. Besides, via a DMLS process, the manufacturability also depends on the cumulative deformation of the layers during the manufacturing process. Existing methods on Design for Additive Manufacturing (DFAM) provide design guidelines that are based on past successful printed designs. However, they are not effective in predicting the manufacturability of metal cellular structures. In this paper, we propose a semi-supervised deep learning based manufacturability assessment (SSDLMA) framework to assess whether a metal cellular structure can be successfully manufactured from a given DMLS process. To enable efficient learning, we represent the complex cellular structures as 3D binary arrays with a simple yet efficient voxelisation method. We then train a deep learning based classifier using only a small amount of experimental data by adopting a semi-supervised learning approach. By running real experiments and comparing with existing DFAM methods and machine learning models, we demonstrate the advantages of the proposed SSDLMA framework. The proposed framework can be extended to predict the manufacturability of various other complex geometries beyond cellular structure in a reliable way even with a small number of training data.

近年来，金属蜂窝结构由于其设计自由度和实现多功能机械性能的能力而受到各个工业领域的关注。但是，即使采用现代增材制造技术（例如直接金属激光烧结（DMLS）工艺），由于其复杂的几何形状，金属蜂窝结构也难以制造。由于结构的几何特征复杂，通过DMLS工艺评估金属蜂窝结构的可制造性是一项艰巨的任务。此外，通过DMLS工艺，可制造性还取决于在制造过程中各层的累积变形。增材制造设计（DFAM）的现有方法提供了基于过去成功的印刷设计的设计准则。然而，它们不能有效地预测金属蜂窝结构的可制造性。在本文中，我们提出了一种基于半监督的基于深度学习的可制造性评估（SSDLMA）框架，以评估是否可以通过给定的DMLS工艺成功地制造金属蜂窝结构。为了实现高效学习，我们使用简单而有效的体素化方法将复杂的细胞结构表示为3D二进制数组。然后，我们采用半监督学习方法，仅使用少量实验数据来训练基于深度学习的分类器。通过运行真实的实验并与现有的DFAM方法和机器学习模型进行比较，我们证明了所提出的SSDLMA框架的优势。