FedScope-KD: Knowledge Distillation-Enhanced Federated Learning via Shared Composition and Personalized Exploration for Heat Emission Prediction in Additive Manufacturing

Abstract Additive manufacturing (AM) enables the production of customized and complex metal parts, providing greater flexibility than traditional mass production. However, maintaining consistent product quality for AM processes remains challenging due to process variability and heat emission fluctuations, which can cause defects such as overheating. Overheating occurs when non-uniform temperature distributes across the printed parts, leading to issues such as structural weaknesses. Predicting heat emissions accurately is essential to preventing such issues, but AM facilities, particularly small- and medium-sized enterprises (SMEs), generate limited and heterogeneous datasets due to unique configurations and privacy concerns, making traditional centralized predictive model training paradigms impractical. Federated learning (FL) offers a privacy-preserving solution, it alleviates data privacy and data scarcity concerns by enabling collaborative learning across decentralized datasets while keeping data local. However, its performance suffers under heterogeneous client data and varied computational resources. This creates a need to address data heterogeneity while maintaining both global model performance and individual client personalization. We propose FedScope-KD, a novel heterogeneity-aware FL framework designed to balance the scope of global and personalized learning. Client heterogeneity level is quantified prior to the FL training, to improve the model weight aggregation during training. This is achieved through component decomposition and knowledge distillation. The framework decomposes client data into process-invariant representations, which capture the fundamental nature of the laser powder bed fusion (LPBF) process, and process-variant representations, which reflect local, client-specific configuration settings. This decomposition ensures that global models benefit from collaborative learning on the homogeneous data representations, while personalized predictions remain tailored to client-specific heterogeneous data representations. A knowledge distillation (KD) mechanism further enhances the framework by efficiently transferring process-invariant knowledge from a global teacher model to local student models. We compare the heterogeneity index generation and calculate the Dynamic Time Warping (DTW) distance for both inter-client variant and invariant components. The variant components exhibit greater distances between clients, reflecting their inherent heterogeneity, while the invariant components are more closely aligned across clients. This pattern demonstrates the effectiveness of the decomposition process in separating heterogeneous and homogeneous information. The heterogeneity index is further used for weight aggregation during FL training. The framework’s performance is validated through a comparison with baseline FL methods such as FedAvg, as well as centralized and individual learning paradigms. Performance metrics include personalized performance and global accuracy with MSE. Through experiments on heat emission datasets in the LPBF process, we show that FedScope-KD successfully addresses the challenges of data heterogeneity and resource constraints, providing a robust solution of predictive modeling for SMEs.