本团队毕业生温红宁博士（香港理工大学）等在Journal of Materials Processing Technology发表研究论文。

摘要：Hot extrusion is a critical manufacturing technology for tailoring microstructure and eliminating metallurgical defects of powder metallurgy (P/M) superalloys in industrial applications. However, the complex coupling between severe plastic deformation and adiabatic heating challenges the prediction and precise control of microstructures. To address this challenge, this study proposed a synergistic end-to-end framework integrating a continuous strain-gradient high-throughput extrusion technique with physics-constrained interpretable machine learning. The established paradigm enables the rapid acquisition of microstructural evolution data under a wide strain range of 0–2, facilitating the determination of grain refinement limit of 3.63 μm at the strain of approximately 1.0. By leveraging the explicit kinetic formulas mined by symbolic regression, we quantitatively decoupled the competing contributions of dynamic recrystallization (DRX) and thermally activated grain growth. The kinetic analysis demonstrates that the grain refinement during the initial stage is primarily governed by the concurrent activation of multiple DRX nucleation ways and dynamic precipitation of primary γ' precipitates at grain boundaries. Then, the adiabatic deformation heat-driven γ' redissolution and subsequent grain growth dominated the later stage after the saturation of DRX. These findings challenge the conventional ‘larger strain, finer grain size’ consensus during extrusion, providing a generalizable optimization strategy for metal extrusion processing that extends beyond the specific case of superalloys.