Given the extensive use of sheet metal-forming processes in the industry and the constant emergence of new materials, the accurate prediction of material constitutive models and their parameters is extremely important to enhance and optimise these processes. Machine learning techniques have proven to be highly promising for predicting these parameters using data obtained either experimentally or through numerical simulations. However, ML models are often constrained by the limited dataset coverage from numerical simulations, which restricts their predictive capability to specific finite element meshes, leading to potential dependency on the discretisation scheme. To address this challenge, a new approach is proposed that integrates ML with inter-extrapolation of strain data to a grid of points within the specimen domain, expanding the dataset coverage and reducing dependency on discrete mesh points. The current work explores this approach by interpolating and extrapolating manipulated data obtained from a Finite Element Analysis, considering a biaxial tensile test on a cruciform-shaped sample. Models are trained and evaluated for performance and robustness. The results show the high accuracy of the interpolated data, along with the excellent performance metrics and robustness of the trained models, ensuring the successful implementation of this approach.

Machine learning models, particularly Extreme Gradient Boosting, have been explored for predicting material parameters in constitutive models that describe the plastic behaviour of metal sheets. While effective for simple constitutive models like Hill′48, their performance declines with more complex models such as the Cazacu-Plunckett-Barlat yield criterion. This study examines the influence of training dataset size and dimensionality reduction via principal component analysis on predictive accuracy. Results show that increasing the training dataset size leads to only marginal improvements, with testing coefficient of determination value plateauing at about 0.50, despite a consistently high training value of about 0.99999, indicating overfitting. Similarly, applying principal component analysis to the baseline model provided no significant enhancement. These findings suggest that simply expanding the dataset or reducing dimensionality is insufficient to address the complexities of CPB′06. Instead, alternative approaches such as advanced feature selection, hybrid ML-physics-based models, or regularization techniques may be required to improve generalization. Future work should explore methods integrating domain knowledge and physics-based modelling to enhance predictive accuracy for complex constitutive models.

Machine learning (ML) algorithms have been studied in literature as an inverse method to predict material constitutive parameters. However, these approaches are often dependent on the mesh discretisation settings applied during numerical simulations, and then difficulty model adaptation to experimental digital image correlation (DIC) subsets. Although a recent study explores the use of an interpolation-based approach to achieve experimental adaptation from numerically-based trained ML models, the proposed methodology lacks evaluation using experimental data. As a follow-up, this study proposes a new evaluation approach. Numerical data is DIC-levelled via MatchID software and then submitted to interpolation. An XGBoost algorithm is then trained on interpolated DIC data and evaluated for parameter prediction, comparing the obtained results with those obtained from the model trained on interpolated numerical data. Overall, the proposed DIC-levelling and interpolation pipeline yields an excellent predictive performance, with results comparable to those obtained when training on interpolated numerical data. The largest deviations are observed for the hardening exponent, while the remaining parameters are predicted with consistently high accuracy. These findings validate the practical applicability of the interpolation-based strategy to reduce the subset scheme dependency of ML models trained on real experimental data.