Scholarlink

<p>Localized pitting corrosion–a long-standing problem for civil infrastructure such as in the hydraulic steel structures motivating this work–involves complex multiphysics processes and exhibits significant stochasticity in both the initiation and propagation phases. Consequently, updating predictive models with measurement data is critical for reliable structural health monitoring of structures subject to such corrosion. Even though previous studies have explored model updating techniques for corrosion prediction, these investigations have been largely limited to proof-of-concept demonstrations or oversimplified models. This study, for the first time, integrates model updating of an advanced hybrid pitting simulation with real-world data generated from accelerated corrosion experiments, utilizing probabilistic machine learning. In the proposed framework, accelerated corrosion experiments are first conducted in accordance with ASTM standards. The resulting pit morphologies are then characterized in detail, with high-resolution laser scanning employed to obtain pit dimension measurements. After that, the pit measurements are used to update a hybrid simulation model developed in our previous study. The variables of interest for updating include: (1) two environmental coefficients of a phase field-based pitting corrosion simulation model and (2) two statistical parameters that controls the pit initiation time. All four parameters cannot be directly observed and need to be updated based on observations. Finally, a summary network and a conditional invertible neural network are employed to efficiently infer the posterior distribution of the four uncertain model parameters based on the accelerated experimental measurements. Results demonstrate that the proposed method achieves strong agreement with independent measurements and outperforms empirical power-law models, especially in predicting upper-percentile pit growth. The improved predictive accuracy validates the promise of probabilistic machine learning approach in correlating the physics-based simulations with real-world experiments. This approach enables more accurate failure prognostics and supports risk-informed optimal maintenance decisions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning-enabled probabilistic hybrid model updating of pitting corrosion with application to hydraulic steel structures: an experimental validation

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Guofeng Qian,
  • Duncan J. Fure,
  • Long Wang,
  • Zhen Hu,
  • Michael D. Todd

摘要

Localized pitting corrosion–a long-standing problem for civil infrastructure such as in the hydraulic steel structures motivating this work–involves complex multiphysics processes and exhibits significant stochasticity in both the initiation and propagation phases. Consequently, updating predictive models with measurement data is critical for reliable structural health monitoring of structures subject to such corrosion. Even though previous studies have explored model updating techniques for corrosion prediction, these investigations have been largely limited to proof-of-concept demonstrations or oversimplified models. This study, for the first time, integrates model updating of an advanced hybrid pitting simulation with real-world data generated from accelerated corrosion experiments, utilizing probabilistic machine learning. In the proposed framework, accelerated corrosion experiments are first conducted in accordance with ASTM standards. The resulting pit morphologies are then characterized in detail, with high-resolution laser scanning employed to obtain pit dimension measurements. After that, the pit measurements are used to update a hybrid simulation model developed in our previous study. The variables of interest for updating include: (1) two environmental coefficients of a phase field-based pitting corrosion simulation model and (2) two statistical parameters that controls the pit initiation time. All four parameters cannot be directly observed and need to be updated based on observations. Finally, a summary network and a conditional invertible neural network are employed to efficiently infer the posterior distribution of the four uncertain model parameters based on the accelerated experimental measurements. Results demonstrate that the proposed method achieves strong agreement with independent measurements and outperforms empirical power-law models, especially in predicting upper-percentile pit growth. The improved predictive accuracy validates the promise of probabilistic machine learning approach in correlating the physics-based simulations with real-world experiments. This approach enables more accurate failure prognostics and supports risk-informed optimal maintenance decisions.