Gas metal arc welding (GMAW)-based additive manufacturing (AM) is a key metal 3D printing technology for the fabrication of near-net shape parts. The thermal history of the multi-layer GMAW-based AM process has a significant influence on part quality and substrate deformations, but it is computationally expensive to accurately calculate it based on numerical method or analytical method. Existing data-driven approaches for the thermal field prediction of AM processes show great advantages in efficiency, but they are mainly applied to single-layer parts or the multi-layer parts with fixed geometry. The main contribution of this work is to realize the thermal field prediction of the multi-layer GMAW-based AM processes with arbitrary geometries by a machine learning approach. Firstly, a novel method for the discretization of deposition process was proposed to make the numerical simulation method for the thermal analysis of AM processes more adaptive and flexible. Then, a unique data structure was developed to extract the deposition state data from the results of the numerical simulation method, which generate the data for the training of the proposed machine learning model. Finally, a physics-based machine learning method based on an ensemble learning model was designed to identify the correlation between the deposition stage and its corresponding thermal field. Validation results showed that the prediction accuracy of the developed method exceeded 94 % when compared with the results of the numerical simulation method, while the time cost of a single prediction process was only at the millisecond level.

基于气体金属电弧焊（GMAW）的增材制造（AM）是用于制造近净形状零件的关键金属3D打印技术。多层基于GMAW的增材制造工艺的热历史对零件质量和基板变形有重大影响，但基于数值方法或分析方法精确地计算它的计算量很大。现有的数据驱动的AM过程热场预测方法在效率上显示出很大的优势，但它们主要应用于具有固定几何形状的单层零件或多层零件。这项工作的主要贡献是通过机器学习方法实现具有任意几何形状的基于多层GMAW的增材制造过程的热场预测。首先，提出了一种离散化沉积过程的新方法，以使AM过程热分析的数值模拟方法更具适应性和灵活性。然后，开发了一种独特的数据结构，以从数值模拟方法的结果中提取沉积状态数据，从而生成用于训练所提出的机器学习模型的数据。最后，设计了一种基于整体学习模型的基于物理学的机器学习方法，以识别沉积阶段与其对应的热场之间的相关性。验证结果表明，与数值模拟方法相比，该方法的预测精度超过94％，而单个预测过程的时间成本仅为毫秒级。