Micro-heterogeneous materials such as dual-phase steels can exhibit variations of the microstructure’s morphology on a length scale, which is not significantly smaller than the considered macroscale. Then, a variation of the effective material properties is induced, which needs to be properly quantified for practical applications. In this contribution, a method based on virtual experiments on a set of artificial microstructures, which follow the real material’s variation, is discussed. From these virtual experiments, characterizing homogenized properties of the material on the macroscale can be derived and the expected variation therein, properly quantified, such that probability density functions or bounding intervals may be used in subsequent uncertainty quantification problems. Since the method itself relies on a Monte Carlo simulation of different artificial microstructures, reducing the required numerical effort is an important challenge. Thus, a machine learning approach will be presented, which is trained to substitute the costly numerical simulations in terms of finite elements by comparatively cheaper models, whilst an acceptable accuracy is maintained. Finally, the set of artificial microstructures is used in a fully-coupled two scale finite element simulation (FE \( ^2\) ), which allows the consideration of locally varying microstructures within a structural engineering problem without the necessity of an a priori computation of relevant material parameters for a classical material model.