Cross-patch graph transformer enforced by contrastive information fusion for energy demand forecasting towards sustainable additive manufacturing

Kang Wang, Haoneng Lin, Naiyu Fang, Jinghua Xu, Shuyou Zhang, Jianrong Tan, Harry Qin, Xuan Liang

Abstract: We propose an effective deep learning method, called Cross-patch Graph Transformer, for predicting the energy demand of Additive Manufacturing (AM) products, which helps to determine the solution with minimal fabrication energy for sustainable AM. This novel method can predict the energy demand of intricate structures by training with simple structures, which alleviates the expensive burden of collecting training data. Our method efficiently integrates node-level, patch-level, and image-level information from part geometry, enabling precise energy demand predictions for products manufactured using AM technology. This approach contributes methodological insights into developing a contrastive information fusion model that enhances energy-related representations even with limited data resources. The incorporation of the cross-patch interaction module enables the method to effectively capture structural relationships, enriching the learning process. Extensive experimental results show our method achieves a higher mean prediction accuracy of 98.3%, validating the effectiveness of our approach across a diverse set of intricate structures. This method not only provides a robust and quantitative tool for identifying optimal solutions with minimal energy demand during the manufacturing of complex structures, but also holds the potential to drive the evolution of computer-aided design towards more sustainable AM practices.