Abstract

Probabilistic emulator models provide a principled approach for uncertainty quantification with expensive computer models. However, modern computer models often produce high-dimensional outputs with relevant modes of variation that are difficult to describe parsimoniously. To address this problem, we combine normalizing flows, a specialized neural network architecture that learns complicated probability density functions via parameterized bijections, with Gaussian processes. The resulting model is more flexible than related methods (e.g., using principal components analysis to represent the outputs) and is shown to better reproduce the computer model outputs. We demonstrate the value of this model for sensitivity analysis and parameter calibration in the context of atomic and plasma physics codes used to simulate laser-induced breakdown spectroscopy.