Data-efficient uncertainty quantification and optimization of electronic systems using machine learning

Paolo Manfredi (Politecnico di Torino)

Abstract: This talk provides an overview over the utilization of Gaussian process regression (GPR) for the data-efficient uncertainty quantification, yield analysis, and optimization of electronic systems. GPR, also known as Kriging, is a nonparametric method belonging to the class of kernel machine learning regressions. As such, it does not assume a predetermined functional form, but it rather constructs the model from data. This makes it more flexible and suitable for building surrogate models of computationally expensive systems characterized by a large number of independent input design parameters. In particular, we will discuss the calculation of statistical moments and the prediction confidence, explore active learning strategies for the iterative acquisition of optimal training datasets, and introduce a hybrid technique that combines GPR and polynomial chaos expansion (PCE).