Data-driven surrogate models for the acceleration of two-scale problems in mechanical engineering
Two- and multi-scale mechanical problems require the homogenization of material behavior at underlying scales for the computation of large scale structures. The formulation of analytic constitutive models identifying all phenomena at small scales is a highly nontrivial problem due to the complexity of materials. A highly efficient alternative is the calibration of machine learning models, e.g., artificial neural networks (ANN), based on experimental or simulation data. The present work illustrates three cases, in which ANN have been successfully trained with simulation data in order to obtain a surrogate model for the effective material law and, therefore, enable the accelerated analysis of large structures with micromechanics-informed material laws. The first case shows for a pseudoplastic three-phase material an on-the-fly model switch between an ANN and a reduced order model, both trained based on RVE computations. The second example illustrates the identification of the temperature-dependent traction-separation law in aluminum grain boundaries based on MD simulation data. Lastly, the final example demonstrates the capability of ANN to approximate the effective material behavior of 3D printed lattice metamaterials with unstable behavior for large deformations.