| Behavioral modeling for analog circuits is in high demand for architectural exploration and system prototyping of increasingly complex electronic systems. The sooner the working system is explored in the overall design process, the clearer the hardware plan for fabrication will be. System exploration using software CAD tools is much less expensive than hardware prototyping. Since SPICE simulations require completely specified circuits to compute responses in the frequency or time domains, simulating the whole system is time consuming, even without considering convergence issues. Behavioral models, which reduce simulation time with acceptable accuracy tradeoffs, are needed in the early stages of system prototyping.;In this research, we focused on developing behavioral-level analog circuit performance modeling methodology using kernel based support vector machine (SVM). In order to improve on model accuracy, we looked at several approaches, including: optimizing learning parameters, sampling schemes, and model selections. An optimizer is developed to tune learning parameters. A sample size investigator is invented to detect minimum number of samples required. A model selector is developed to automatically select the model obtaining the highest accuracy. With all three components combined into a modeling flow, accuracy is greatly improved. However, the success of these types of techniques depends on large sets of training data. And analog data is expensive in terms of simulation time and hardware testing. Therefore, achieving high modeling accuracy with limited datasets has become a challenge. Therefore, we further developed a sampling method dynamically forms datasets based on its selection of dominant support vectors, requiring less data while maintaining almost the same level of model accuracy.;In terms of accuracy of the model, we computed average error, worst case error, and special design corner case error, by compared against SPICE simulation result. We found because of the nature of SVM, it trades off corner case accuracy for learning the general behavior trend. Therefore, for improving worst case and corner case accuracy, we applied Gaussian process for machine learning approach, and results show great accuracy improvement with reasonable evaluation speed tradeoffs. |
