A preliminary study of different metrics for the validation of device and behavioral models

In integrated circuit and/or system design using CAD tools, model accuracy has obviously a direct impact on the quality of the final performance prediction. One of the most common ways to evaluate the quality of a model has always been the visual inspection of the comparison between simulations and measurements; the highest quality is assigned to the model that returns the set of curves closer to the measured ones. Unfortunately a large number of variables (bias, frequency, power level, large signal regime, etc.) may be of concern and, in such a case, it can be quite hard to decide which is the best model for a targeted application. Therefore, the problem of accuracy evaluation can be addressed by defining suitable "metrics", i.e., numerical formulae which return, possibly, dimensionless quantity as performance indexes for a model in term of best fitting between simulations and measurements. In this work several metrics have been studied from the model validation point of view. In particular, different metrics for DC characteristics, scattering parameters as well as large signal performance have been tested for HBT and NMOS transistor models, and for WCDMA base station LDMOS power amplifier models. Dis/Advantages of the proposed metrics and possible solutions/improvements are discussed.