本团队博士生余国卿等在International Journal of Plasticity发表研究论文。

摘要：Thermomechanically processed metallic alloys usually exhibit complex anisotropic yield behavior induced by crystallographic texture. Owing to their fixed mathematical form, conventional yield criteria possess limited accuracy and flexibility in capturing such anisotropy, and they require costly parameter recalibration when applied to different materials or crystallographic textures. This paper introduces the generalized convex yield network (GCYN), a feedforward neural network with partial input convexity constraints to predict the yield onset. The proposed architecture decomposes the yield function into a convex stress-dependent branch that guarantees global convexity of the yield surface, and a non-convex branch integrates tailored material descriptors, such as generalized spherical harmonics (GSH) coefficients and typical yield onsets, to explicitly capture anisotropic yield behavior across distinct textures and materials. This bifurcated design enables the GCYN to learn texture/material-specific corrections to the yield surface. The generalizability of GCYN across textures was demonstrated by training on yield datasets for different textured TA1 titanium generated via crystal plasticity (CP) simulations. The well-tuned criteria accurately predicts the yield onsets when applying to the untrained textures, achieving an average prediction error of 0.903% in the 6D stress space. In addition, GCYN accurately predicted the yield onsets of some experimental titanium textures with different characteristics. To demonstrate the potential applicability of GCYN to different materials at a conceptual level, the model was trained on datasets generated from the Lode-dependent anisotropic-asymmetric (LAA) Hill48 criterion. The trained model accurately emulates a parameterized family of analytical yield loci, achieving an average prediction error below 0.2% in plane. Moreover, The yield loci predicted by the GCYN strictly satisfy global convexity while preserving the essential characteristics of the original yield surface. This work provides a fully convex and material/texture-sensitive yield criterion.