We discuss a Deep Learning (DL) approach for the multi-scale characterization of polysilicon movable structures of MEMS (micro-electro-mechanical systems), based on data assimilation from two-dimensional stochastically representative images of polycrystalline films. A dataset of microstructures is collected and a convolutional neural network is trained, to provide the appropriate scattering in the value of the overall stiffness (in terms e.g. of Young’s modulus) of the grain aggregate. Since results are framed within a stochastic approach, the aim of the learning stage is to provide a fast method to be next used at the device level for statistical, Monte Carlo-like investigations of the relevant performance indices. Accuracy of the proposed approach is assessed for different values of the ratio between the dimension of the polycrystalline aggregate and the representative size of a single grain, i.e. for different number of grains gathered in the polycrystal, to next check if size effects are correctly captured.