A Hybrid Surrogate Modeling Method for Corrosion Morphology Prediction Under Non-Stationary Dynamic Loading

Abstract

The loads on the structures or components are more often fluctuating stochastically over time and exhibit non-stationary statistical behaviors. The goal of this research is to predict the corrosion growth of large civil infrastructure using multi-scale simulations with the consideration of time-dependent stochastic loads along with other sources of uncertainty. A stochastic process modeling method is first employed to model the stochasticity of the load conditions based on real-world data. A multi-scale simulation model is then constructed to predict the corrosion morphology evoluation over time subject to dynamic loads. The model consists of a mesoscale phase-field simulation and a macroscale structural analysis model. The required high computational effort of the multi-scale simulation model makes it unsuitable for probabilistic analysis purposes, which requires the execution of the model thousands of times. This research proposes a hybrid surrogate modeling to substitute the original multi-scale simulation model. The hybrid surrogate model consists of a deep neural network-based autoencoder for dimension reduction and a Gaussian process regression-based model for forecasting and uncertainty quantification. A navigational lock miter gate case study is employed to demonstrate the efficacy of the proposed method.