As the push for higher operating frequencies continues, especially in the millimeter-wave (mmw) range, ensuring optimal electromagnetic (EM) performance of the system being designed has become increasingly challenging. At these frequencies, even minor manufacturing variations in geometric parameters at package and PCB levels can significantly affect RF performance, leading to poor yields or limiting system capabilities. Traditionally, process-voltage-temperature (PVT) corner analyses are widely applied at the transistor and circuit levels to ensure robustness across manufacturing variations and operating conditions. However, systematic analysis of process variations, due to manufacturing tolerances, is rarely extended to the package and PCB domains, although their impact on high-frequency signals is large. This research aims to bridge this gap by introducing a structured approach to model parameterized Die+package+PCB RF transitions using a combination of electromagnetic simulations and machine learning (ML) techniques, potentially allowing cheaper package technologies to be exploited also for millimeter-wave designs. A parameterized antenna-in-package model was developed in Ansys HFSS 3D Layout, and space-filling Latin Hypercube Sampling was used to explore the design space efficiently. Surrogate models were trained in OptiSLang and compared, after which a Gaussian Process framework in GPyTorch was implemented to predict S-parameters and antenna gain with high accuracy. The results show that the surrogate models reproduce full-wave simulations with low error while reducing computational cost by orders of magnitude. This enables fast prediction of EM performance under process variation and systematic extraction of EM corners. The work demonstrates a scalable flow that integrates EM simulation and machine learning, providing an effective tool for robust package-level design.