accurate description of sheet metal plasticity [1], but their increased parameterization and nonlinear structure pose significant challenges for data-driven parameter identification. While machine learning (ML) techniques, particularly tree-based methods such as XGBoost, show excellent predictive performance for simpler criteria (e.g. Hill’48), their generalization capability degrades significantly when applied to CPB’06, even with large synthetic datasets [2]. This work investigates strategies to improve the predictive quality of ML-based calibration for the CPB’06 yield criterion. Using numerically generated data from biaxial tensile tests on cruciform specimens, the study examines the relationship between constitutive model complexity and learning performance, focusing on feature relevance, parameter sensitivity, and training strategy design. Instead of relying solely on larger datasets or generic dimensionality reduction, the approach emphasizes the identification of weakly influential parameters and the incorporation of physics-informed constraints. Preliminary results show that better alignment between the ML formulation and the underlying mechanics can reduce overfitting and improve prediction robustness. The findings indicate that CPB’06 parameter identification is primarily a modeling challenge rather than a data-volume issue, highlighting the need for tighter coupling between constitutive theory and machine learning.