Methods of Machine Learning for Analysis and Decoupling of Power Delivery Networks on Printed Circuit Boards

Increasing challenges in the design process of power delivery networks on printed circuit boards require new and adapted tools. Therefore, in this thesis artificial neural networks are investigated to enhance the design process. Challenges and their possible solutions are discussed. For the data driven investigations more than 100000 numerical simulations of printed circuit board variations using a physics-based modeling approach are performed. Additionally, more than 124000 decoupling capacitor (decap) terminations on the power delivery networks are processed. Validations of some printed circuit board variations against a ommercial full-wave finite element method solver are included. The investigations resulted in the development of the publicly available SI/PI-Database. The database holds a majority of the printed circuit board descriptions and simulation results. It helps to increase the reusability of once created data samples for different investigations throughout this and future work. Here, the database is used in combination with a fine-tuning approach of the artificial neural network training process to increase the data efficiency of once created data samples. In the fine-tuning process the existing data samples are used to better initialize the artificial neural network for the training.