Sie sind hier: Startseite Kolloquium 2022 Physics-guided neural networks in solid mechanics

Physics-guided neural networks in solid mechanics

Prof. Dr. Oliver Weeger (TU Darmstadt)

Art des Termins
    Wann 10.02.2022
    von 10:00 bis 11:00
    Wo Zoom
    Termin übernehmen vCal

    Machine learning methods such as artificial neural networks are increasingly gaining attention for applications in solid mechanics and dynamics. They can be beneficial when high-fidelity models must be reduced to enable real-time simulations, e.g., for digital twins and co-simulations, or when complex models cannot be represented analytically, e.g., in constitutive relations for multiscale and multiphysics simulations. However, for machine learning-based models to become as reliable and robust as conventional ones, it is crucial to comply to physical and mathematical requirements as much as possible. This can be achieved by incorporating such information directly into the formulation, i.e., develop a physics-guided model, or by using training data that includes all relevant requirements, i.e., make the model physics-informed.

    In this talk, we will first introduce physics-guided constitutive models for material modeling and multiscale simulations. Using input convex neural networks, we develop high-flexible hyperelastic, visco-hyperelastic, and electro-mechanical constitutive models that incorporate all physical requirements such as thermodynamic consistency, objectivity, material symmetry, material stability, and growth-conditions into the model formulations. Furthermore, we discuss the development of surrogate models for parametrized dynamical systems, which are formulated as neural ordinary differential equations.