Training deep material networks to reproduce creep loading of short fiber-reinforced thermoplastics with an inelastically-informed strategy

Abstract Deep material networks (DMNs) are a recent multiscale technology which enable running concurrent simulations on industrial scale with the help of powerful surrogate models for micromechanical problem. Classically, parameters DMNs identified based linear elastic precomputations. Once identified, may process inelastic and were shown to reproduce full-field original microstructure high accuracy. The work at hand was motivated by creep loading thermoplastic components fiber reinforcement. In this context, multiple scales appear, both in space (due reinforcements) time (short- long-term effects). We demonstrate computational examples that classical training strategy precomputations is not guaranteed produce whose response accurately matches high-fidelity computations. As remedy, we propose an inelastically informed early stopping offline DMNs. Moreover, introduce novel model, shares principal nonlinear effects true model but significantly less expensive evaluate. For problem hand, enables saving significant during parameter identification process. provides reliably generalize loading.