Conventional Physics-Informed Neural Networks (PINNs) for Metal Injection Molding (MIM) sintering often encounter thermodynamic inconsistencies, such as non-physical density regression during the cooling stage. This study presents a Dual-Stream Decoupled PINN (DSD-PINN) that incorporates a novel time-freezing gating mechanism to enforce thermodynamic irreversibility. A multi-stage kinetics model is integrated to capture the complex densification behavior of 17-4PH stainless steel. Validated against industrial-grade ANSYS simulations, the DSD-PINN effectively mitigates the density rollback observed in standard PINN variants. Quantitatively, the framework maintains a relative L 2 -norm prediction error of 7.25×10⁻² (below 7.3%) and achieves a 267× improvement in online inference efficiency. By reinforcing physical consistency through its architecture, this model provides a reliable computational engine for high-frequency digital twins and real-time process monitoring in industrial MIM systems.
Bai et al. (Wed,) studied this question.