On high-accuracy global corrosion parameter determination in metallic plates using multi-random-field inverse surrogates

Abstract

The accurate assessment of global corrosion characteristics in metal plates, particularly thickness loss and surface roughness, is critical for industries such as oil, marine, and energy. Recent advances using guided ultrasonic waves have enabled non-destructive evaluation of corroded structures; however, this task remains challenging due to complex interactions between wave propagation and spatially varying thickness profiles. The difficulty is further compounded by the inherently heterogeneous nature of real corrosion patterns, which introduces significant uncertainty in signal interpretation and limits the effectiveness of traditional physics-based identification methods. In this study, global corrosion parameters, including mean thickness reduction and thickness standard deviation, are reliably estimated using data-driven inverse surrogate models. Corrosion morphology is modeled as a stochastic field, while guided wave responses are generated via numerical simulations. The inverse framework maps features extracted from wave responses to corresponding corrosion parameters. A key challenge lies in the pronounced non-uniqueness of the problem, as infinitely many stochastic field realizations can yield identical statistical descriptors. To address this issue, a multi-random-field strategy is introduced, in which guided wave responses from multiple independent realizations are averaged and jointly analyzed. Extensive comparative studies show a clear improvement in prediction accuracy with as few as three random field realizations, with further significant gains observed for ten random field configurations. For the considered test cases, the relative mean absolute error was reduced to approximately 1.4% for thickness reduction estimation and 7.0% for standard deviation estimation. Across all cases, the proposed inverse modeling approach outperforms conventional time-of-flight-based identification methods derived from dispersion curve analysis. Although this research focuses on computational modeling, in practical applications, the proposed methodology can be readily implemented by repositioning the sensor network or excitation source, either laterally or rotationally.