Pitting Corrosion Model Updating Using Experimental Data

Abstract

Both physics-based (deterministic) and statistical models are widely used in structural health monitoring (SHM) research due to their ability to predict damage progression, particularly in scenarios with limited data availability. However, while all parameters in statistical models and certain parameters in physics-based models cannot be directly measured, updating these parameters using measurement data is essential for ensuring accuracy. This study presents a novel approach to updating a hybrid corrosion model using accelerated corrosion experimental data combined with an amortized likelihood-free Bayesian inference method. In the proposed framework, pit numbers and dimensions developed after different corrosion time durations are first measured using a high-precision laser scanner. Two reaction coefficients and two statistical parameters, which are not directly measurable, are selected for updating in the hybrid simulation. A summary network and a conditional invertible inference network are jointly trained on extensive simulation data to learn the mapping between observed data and model parameters. The trained network is then utilized to infer the posterior distribution of the model parameters based on experimental measurements. The results demonstrate that the proposed conditional invertible neural network-based inference method can effectively and efficiently update the hybrid model, enhancing parameter estimation for complex damage mechanisms such as corrosion. Additionally, a key advantage of this approach is its computational efficiency, as it can perform inference within seconds—significantly faster than other likelihood-free inference methods. This capability makes it particularly suitable for real-time SHM applications, improving both accuracy and practicality in damage assessment.