Mechanical structures are made of materials with complex behaviors. Accurate structural simulations depend on integrating physically consistent constitutive models, typically calibrated using experimental data and inverse methods. Traditionally, this calibration involves selecting an analytical constitutive law and fitting its parameters to experimental data. Although interpretable, such models often struggle to accurately represent complex material responses. To overcome this limitation, data-driven approaches have emerged, directly identifying constitutive laws from data without predefined assumptions [1,2]. Some data-driven methods embed thermodynamic consistency by leveraging convex energy potentials to ensure physical reliability. However, these approaches commonly assume predefined forms for the potentials and internal variables, restricting their applicability when such prior knowledge is unavailable. To overcome this limitation, we propose a neural network-based method [1] that simultaneously identifies energy potentials and internal variables without predefined structural assumptions [1,2]. The approach maintains thermodynamic consistency and utilizes the Constitutive Relation Error (CRE)—an indicator providing physically meaningful error measurements based on full-field data—and its modified version (mCRE), accounting for measurement noise. We first apply our approach to identify the viscoelastic behavior of materials used in vascular surgical simulators. Experimental data are acquired using Digital Image Correlation during uniaxial and biaxial tests. Our objective is to extend the mCRE framework coupled with neural networks [1] to learn the underlying thermodynamic potentials governing the constitutive law [1,2]. To validate the method, we also study a controlled academic example: a plate undergoing elasto-visco-plastic deformation described by the Marquis–Chaboche model. The analytical formulation developed by Ladevèze et al. [3] serves as a reference, enabling to assess the framework's capability to recover known constitutive components. These studies demonstrate the robustness of the studied framework, combining physics-based modeling, meaningful error indicators, and unsupervised identification of constitutive laws. References : [1] Rosenkranz et al. (2024). Comput. Mech., 74(6), 1279-1301. [2] Benady et al. (2024). Int. J. Numer. Methods Eng., 125(8), e7439. [3] Ladevèze & Chamoin (2024). Adv. Model. Simul. Eng. Sci., 11(1), 23.