Antenna tuner design flow optimization

In this thesis an approach to an optimization of the antenna tuner design flow is examined. In the antenna tuner design process, numerous design iterations are required to fulfill a given design specification. This usually requires EM-simulations that increase in time with the number of tuners on the chip and can reach hours of simulation time. An analytical model for the maximum voltage handling capability and the capacitance in the off-state of stacked RF-switches is difficult to achieve, as the capacitance and voltage handling are dependent on parasitic influences in layout. Further, to achieve good voltage handling, a ballasting algorithm is used for designing antenna tuners, which in turn is dependent on the resulting scatter parameters from layout EM simulations, and influences the off-state capacitance. To overcome this, a machine learning model, based on layout and circuit parameters, is investigated for the off-state capacitance of the antenna tuner and the maximum voltage handling capability after ballasting, and its prediction performance is analyzed. Numerous machine learning models, such as neural networks, gradient boosting, ridge regression, and support vector machines, were explored to achieve good accuracy. This includes the analysis of fitting machine learning features that are needed to describe an accurate relation of layout and circuit parameters with the required performance parameter targets. An accuracy of 94.69% and 98.62% for the prediction of the off-state capacitance and maximum voltage handling capability was achieved. As the starting point of designing an antenna tuner is the specification, the machine learning models are used as a surrogate model for a design optimization approach that recommends a design with given specifications. A brute-force solution and an iterative optimization, based on genetic algorithms, are presented and compared.