MATSE—multi-fidelity augmented time-series emulation: galvanic corrosion applications

Abstract Complex physio-chemical phenomena can be investigated using high-fidelity physics-based finite element (FE) models. These models provide accurate predictions of the response of interest, as well as insights into its evolution over time. However, these models are computationally expensive, limiting widespread use. Machine learning (ML)-based protocols can significantly accelerate predictions from high-fidelity models. However, the training costs associated with these models can be prohibitive. This expense can be mitigated by effectively integrating results from lower-fidelity models. Therefore, a protocol that leverages low-fidelity results to accelerate high-fidelity predictions is desirable. This work introduces a multi-fidelity Gaussian process regression framework designed to emulate time-series outputs from high-fidelity models. The framework integrates low-fidelity model responses with Taylor-series approximations, facilitated by automatic differentiation, to propagate predictions and uncertainties across time and fidelity. The protocol is validated by accurately predicting both analytical and FE examples. A well-known example for time-series models, specifically the pendulum with moving support point and the evolution of the total current output of corroding galvanic couples in two case studies across five model fidelities are investigated. Although developed in the context of corrosion, this versatile ML framework shows immense potential for various engineering applications.