Artificial neural network to statistically model the variation in small signal equivalent circuit model parameters for a Si/SiGe HBT process

On a single wafer and between different wafers, there are variations in device characteristics due to processing non-uniformity and non-reproducibility. In order to capture these variations, we could extract a complete equivalent circuit model for each device. Since this is a time-consuming process, we developed an alternative approach based on an Artificial Neural Network (ANN). This ANN takes measured quantities as inputs, and generates the model parameters as outputs. To keep the complexity of the ANN down to a reasonable size, we limited this mapping to the most sensitive elements. These are determined by performing a sensitivity analysis on a reference device. To demonstrate the method, we applied it to a Si/SiCe HBT process. Results show that the ANN very well predicts model parameters and that it is B very good statistical model.