Abstract
Moving heat source problems are commonly seen in many manufacturing applications, such as machining, laser cutting, welding, and additive manufacturing processes, while numerical modeling often takes time to analyze. A data-driven framework for predicting the spatial-temporal temperature response of a moving heat source has been developed to address the need for real-time temperature prediction of a moving heat source. This paper presents a method to reduce the training data requirements and improve the overall efficiency and accuracy of such a framework. This process, referred to as the boundary-focused training approach, is a type of active learning approach which allows for further scalability of the framework to large and/or complicated geometries by using a clustering approach to classify the spatial data according to differences in its thermal response. Then, a tandem learning approach is adopted to train a series of neural networks for each respective cluster. This boundary-focused training approach is implemented on two materials with different thermal diffusivities to evaluate the suitability of this framework in various contexts. The boundary-focused method showed improvement on the boundary-adjacent cases.
