This paper presents a novel solution to predict the visual defects, one of the major criteria for analyzing manufacturability for the Laser-based Powder Bed Fusion (LPBF) process. For the existing manufacturability investigations, the key challenge is how to model the complex interrelationships among the design, process, and final quality. The recent research proposed machine learning methods to model the manufacturability analysis. Voxel-based Convolutional Neural Network (CNN) has been investigated as one potential solution for design shape analysis. Those approaches are limited by the computational capability available and only lower resolution was performed. However, low resolution is not enough for analyzing the LPBF manufacturing process precisely. Some detailed features may be omitted through the voxelization process. To solve this issue, a more efficient CNN is proposed in this paper. Design data is stored in a sparse matrix so that only the occupied voxels are trained by CNN operations. It joins with the process data, which is trained by a Neural Network (NN) model to make the prediction of manufacturability. By performing the generalized convolutions, the computational costs decrease sharply compared to voxel-based CNN, which offers the advantage of performing with high resolutions. The approach is validated in terms of effectiveness and efficiency on the manufacturability prediction for the LPBF process.

本文提出了一种预测视觉缺陷的新解决方案，这是分析基于激光的粉末床融合 (LPBF) 工艺可制造性的主要标准之一。对于现有的可制造性调查，关键挑战是如何对设计、过程和最终质量之间复杂的相互关系进行建模。最近的研究提出了机器学习方法来模拟可制造性分析。基于体素的卷积神经网络 (CNN) 已被研究作为设计形状分析的一种潜在解决方案。这些方法受到可用计算能力的限制，只能执行较低的分辨率。然而，低分辨率不足以精确分析 LPBF 制造过程。通过体素化过程可能会省略一些详细的特征。为了解决这个问题，本文提出了一种更高效的 CNN。设计数据存储在一个稀疏矩阵中，因此只有占用的体素才能通过 CNN 操作进行训练。它与经过神经网络 (NN) 模型训练的过程数据相结合，以预测可制造性。通过执行广义卷积，与基于体素的 CNN 相比，计算成本急剧下降，后者提供了以高分辨率执行的优势。该方法在 LPBF 过程的可制造性预测的有效性和效率方面得到验证。通过执行广义卷积，与基于体素的 CNN 相比，计算成本急剧下降，后者提供了以高分辨率执行的优势。该方法在 LPBF 过程的可制造性预测的有效性和效率方面得到验证。通过执行广义卷积，与基于体素的 CNN 相比，计算成本急剧下降，后者提供了以高分辨率执行的优势。该方法在 LPBF 过程的可制造性预测的有效性和效率方面得到验证。