Efficient generation of data sets for one-shot statistical calibration of RF/mm-wave circuits

Millimeter-wave circuits in current nanometric technologies are especially sensitive to process variations, which can seriously degrade the device behavior and reduce fabrication yield. To tackle this issue, conservative designs and large design margins are widely used solutions. Another approach consists in introducing variable elements, also called tuning knobs, to allow post-fabrication tuning. One-shot statistical calibration techniques take advantage of advanced machine learning regression tools to propose a set of tuning knobs values that enhance the circuit performance based on simple measurements. Training the regression models require a huge amount of data covering the device performances, the effect of the tuning knobs and the simple measurements that guide the regression. In this work, we propose an efficient method for generating such a data set that reduces noticeably the size of the required training set for an accurate calibration.