A Physics-Informed Residual Learning Framework for Predicting Sintered Density of Binder Jetting Inconel 625

Abstract The expansive and highly coupled process space inherent in binder jetting additive manufacturing presents significant challenges to purely data-driven machine learning models, which often lack the physical interpretability and generalizability required for reliable process optimization and high-quality fabrication. This study proposes a hybrid modeling framework PINN-XGBoost for accurately predicting the sintered density of binder-jetted Inconel 625 by integrating physics-informed neural networks with residual learning based on XGBoost. A comprehensive experimental investigation was conducted to evaluate the effects of binder saturation, layer thickness, and sintering temperature on green part quality, densification behavior, and microstructural evolution. The results reveal strongly coupled relationships among process parameters, with sintering temperature being the dominant factor, while binder saturation and layer thickness affect densification through their influence on green part formation. By embedding physical constraints into the learning process and correcting residuals using XGBoost, the proposed framework demonstrates improved predictive accuracy and generalization over conventional data-driven models. This approach provides a reliable and interpretable tool to support process optimization and promote the practical application of binder jetting additive manufacturing for nickel-based superalloys.