Dynamic physics‐guided neural network for predicting hot deformation behavior of TiAl‐based intermetallic alloys

Abstract Ti‐Al‐based intermetallic compounds are promising candidates for high‐temperature structural applications owing to their outstanding mechanical properties. Ti 2 AlNb alloys, characterized by complex multiphase microstructures, present significant challenges for hot deformation modeling because of their atypical flow behavior and sensitivity to processing conditions. In this study, we systematically investigated the hot deformation behavior of Ti 2 AlNb through experiments and compared conventional constitutive models with advanced machine learning approaches. The conventional strain‐compensated Sellars (SCS) model showed limited accuracy for Ti 2 AlNb, especially across complex microstructural transitions, while performing well for simpler alloy systems like Ti4822. To address these limitations, we developed a dynamic physics‐guided neural network (DPGNN) that integrates physical constraints with data‐driven learning via an adaptive gating mechanism. The DPGNN model significantly outperformed the SCS model and three purely data‐driven baselines, achieving high accuracy (test R 2 > 0.98) and robust generalization across both Ti 2 AlNb and Ti4822 alloys. These findings highlight the value of embedding physical principles within machine learning frameworks, providing a robust and generalizable tool for predicting hot deformation behavior in advanced alloys.