A Bayesian Augmented-Learning framework for spectral uncertainty quantification of incomplete records of stochastic processes

A novel Bayesian Augmented-Learning framework, quantifying the uncertainty of spectral representations stochastic processes in presence missing data, is developed. The approach combines additional information (prior domain knowledge) physical with real, yet incomplete, observations. deep learning models are trained to learn underlying process, probabilistically capturing temporal dynamics, from physics-based pre-simulated data. An ensemble time reconstructions provided through recurrent computations using learned models. Models characterized by posterior distribution model parameters, whereby uncertainties over models, and all quantified. In particular, three neural network architectures, (namely long short-term memory, or LSTM, LSTM-Autoencoder, LSTM-Autoencoder teacher forcing mechanism), which implemented a framework variational inference, investigated compared under many data scenarios. example dynamics pertaining characterization earthquake-induced excitations even when source load records incomplete used illustrate framework. Results highlight superiority proposed approach, adopts information, versatility outputting forms results probabilistic manner.