Séminaire Clément Jailin

On the learning of hyperelasticity

Physics-Augmented Neural Networks (PANNs) have emerged in the numerical community as a powerful approach for learning constitutive behavior. By embedding thermodynamic principles and convexity constraints within the neural network architectures, PANNs satisfy the flexibility of data-driven modeling with the robustness of physics-based formulations. While most developments remain numerical, I will present their transfer to experimental learning using the EUCLID framework, which exploits equilibrium conditions and full-field displacement/force data. Applied to 3D printed TPU samples, this approach enables the discovery of hyperelastic behavior beyond standard models such as Neo-Hookean or Saint-Venant–Kirchhoff laws. Finally, I will discuss extensions toward multi-material learning with Material-Embedding PANNs, which aim to capture entire classes of constitutive behaviors rather than a single material response.