Recent silicon technologies are especially prone to imperfections during the fabrication of the circuits. Process variations can induce a noticeable performance shift, especially for high frequency devices. In this thesis we present several contributions to tackle the cost and complexity associated with testing mm-wave ICs. In this sense, we have focused on two main topics: a) non-intrusive machine learning indirect test and b) one-shot calibration. We have in particular developed a generic method to implement a non-intrusive machine learning indirect test based on process variation sensors. The method is aimed at being as automated as possible and can be applied to virtually any mm-wave circuit. It leverages the Monte Carlo models of the design kit and the BEOL variability information to propose a set of non-intrusive sensors. Low frequency measurements can be performed on these sensors to extract signatures that provide relevant information about the process quality, and consequently about the device performance. The method is supported by experimental results in a set of 65 GHz PAs designed in a 55 nm technology from STMicroelectronics. To further tackle the performance degradation induced by process variations, we have also focused on the implementation of a one-shot calibration procedure. In this line, we have presented a two-stage 60 GHz PA with one-shot calibration capability. The proposed calibration takes advantage of a novel tuning knob, implemented as a variable decoupling cell. Non-intrusive process monitors, placed within the empty spaces of the circuit, are used for predicting the best tuning knob configuration based on a machine learning regression model. The feasibility and performance of the proposed calibration strategy have been validated in simulation.