Stacked Sparse Autoencoders and Classical Artificial Neural Networks for the Inverse Uncertainty Quantification of Dynamic Engineering Systems Models

The paper proposes an integrated framework for Inverse Uncertainty Quantification (IUQ), based on the effective combination of: Stacked Sparse Autoencoders for the dimensionality reduction of time dependent, scarce data; Artificial Neural Networks metamodels, to reduce the computational burden typically associated to system simulation codes; Genetic Algorithms to solve the inverse optimization problems related to the IUQ process; and Sliced Normal distributions to describe complex, asymmetric and multi-modal dependencies. The approach is shown to perform satisfactorily on the 2020 NASA UQ Challenge: accurate and robust characterizations of uncertainty (together with model discrepancies) are produced, while reducing the overall computational cost by two orders of magnitude.