The simulation of large-area diffractive optical elements (DOEs) is challenging when non-paraxial propagation and coupling effects between neighboring structures shall be considered. We developed a novel method for the farfield simulation of DOEs, especially computer-generated holograms (CGHs) with lateral feature sizes in the wavelength range. It uses a machine learning approach to predict the optical function based on geometry parameters. Therefore, training data are calculated by physical simulation methods to create a linear regression model. With the trained model a very fast computation of the farfield is possible with high conformity to results of rigorous methods.