Inelastic Constitutive Kolmogorov-Arnold Networks: A Generalized Framework for Automated Discovery of Interpretable Inelastic Material Models
Please login to view abstract download link
The rapid progress of automated and interpretable material model discovery using machine learning has significantly advanced data-driven constitutive modeling. In this context, Kolmogorov–Arnold Networks (KANs) provide substantially enhanced transparency compared to conventional multilayer perceptrons and have recently demonstrated strong potential in computational mechanics, including the discovery of hyperelastic material models. Building on a generalized constitutive framework for inelastic materials, this talk introduces inelastic Constitutive Kolmogorov–Arnold Networks (iCKANs). As visualized in Figure 1, this formulation combines the flexibility of data-driven modeling with the intrinsic interpretability of KANs, enabling the identification of human-readable inelastic constitutive laws. To explicitly recover interpretable material representations, symbolic regression techniques are employed to express both the strain energy function and the dissipation potential in closed mathematical form. The proposed framework is evaluated using synthetic data as well as experimental viscoelastic datasets of the polymeric materials VHB 4910 and VHB 4905. The results demonstrate that iCKANs accurately capture complex viscoelastic behavior while preserving physical interpretability. Moreover, the framework naturally accommodates additional material features, such as temperature dependence, highlighting its potential as a transparent and extensible approach for data-driven inelastic material modeling.
