Data-driven material modeling
D4
The project D4 – Data-driven material modeling – investigates how accurate and flexible constitutive models for complex materials can be developed using only experimentally measurable data.
The focus lies on nonlinear material behavior, such as anisotropic hyperelasticity, where classical model calibration is often constrained by restrictive assumptions and the need for extensive a priori knowledge. To overcome these limitations, the project follows a consistent two-step data-driven approach: First, data-driven identification (DDI) is used to determine tuples consisting of stress and strain states from full-field displacement measurements, such as those obtained by digital image correlation (DIC). This method allows identifying material response from measurable boundary conditions and deformation fields without assuming a predefined constitutive law.
In the second step, the identified data are used to train a physics-augmented neural network (PANN). By integrating physically motivated constraints – such as material frame indifference and thermodynamic consistency – into a flexible neural network architecture, the model not only captures known material behavior accurately but also allows for reliable extrapolation to previously unseen, but physically plausible, material states.
To evaluate the performance and generalization ability of this framework, benchmark studies are conducted based on synthetic data that replicate realistic experimental conditions. The overarching goal is to develop a robust and fully data-driven alternative to classical model calibration – suitable for capturing the complex behavior of heterogeneous materials and supporting more accurate, application-oriented simulations of modern metamaterials.