We present a methodology to systematically extract efficient and physically-based reduced-order models based on a mixed-level simulation approach and demonstrate its practicality for the design of a capacitive MEMS microphone. The method has been implemented in a MATLAB toolbox. Starting from a FEM discretization, it enables the automated generation of mixed level VHDL-AMS based macromodels, which can be straightforwardly fed into a standard circuit simulator. The whole model generation process is described in all details. The resulting macromodels are highly efficient and moreover, in contrast to other equivalent network approaches, physically-based. Therefore they allow for the predictive simulation of microstructures with complex geometry, as it is demonstrated by the numerical analysis of a capacitive microphone.