Enclosing the modeling error in analog behavioral models using neural networks and affine arithmetic

One all-time challenge in behavioral modeling is to minimize the modeling error while still profiting from a simplified representation of an analog circuit. In many cases the modeling error is known, but up to now it was only an indicator for the quality of the model. Its influence on errors during simulation could not be evaluated. We present a flow for the generation of behavioral models based on neural networks which uses affine arithmetic to guarantee enclosing the modeling error. We also demonstrate that the approach can also be applied to modeling the effects of parameter deviations.