Abstract
Moving heat source problems are prevalent in manufacturing. While numerical models provide comprehensive, multi-scale analyses, they can be time-consuming. This paper introduces a convolutional neural network (CNN)-based framework designed to rapidly predict temperature distributions and presents two methods to enhance efficiency and accuracy when scaling the framework to large 3D geometries. The first method, termed geometric subsection training, reduces the amount of spatial data needed by over 90% for the specific 3D geometry used in this framework. The second method, referred to as the boundary focused training method allows for further scaling the framework to large and complicated geometries by using a clustering approach. Then, a tandem learning approach is adopted to train a series of neural networks for each respective cluster. These methods are implemented on a complex 3D geometry with a random sequential moving heat source as a proof of concept. Results show a high level of agreement with the ground truth generated by finite element analysis. The scalability and limitations of this approach are also discussed in this paper.
