*Result*: A hybrid deep learning model driven by physical mechanisms and data for predicting corrosion in natural gas pipelines.

*In order to address the challenges posed by complex feature correlations, high uncertainty, and insufficient model generalization in predicting the corrosion depth of natural gas pipelines under small sample conditions, this paper proposes a hybrid deep learning framework that integrates physical mechanisms with data-driven approaches. The framework utilizes a Bayesian Network (BN) to identify seven critical features and constructs six interactive features based on physical-electrochemical corrosion mechanisms to enhance physical consistency. The model employs a three-stage architecture: XGBoost serves as the baseline model to learn global nonlinear trends and generate initial predictions. The Kolmogorov-Arnold Network (KAN) is first embedded to perform high-order feature modeling on the residuals of corrosion predictions, enhancing stable representation capabilities. The Gaussian Process (GP) performs residual smoothing correction in the embedded space and outputs a 95 % confidence interval. Validation based on 242 sets of sample data collected from excavation sites of buried pipelines in southern Mexico that have been in service for over 50 years.The findings indicate that by employing Bayesian methods for joint hyperparameter adjustment, the model attains a prediction performance of R2 = 0.9613 and a root mean square error (RMSE) of 0.2809 on a dataset comprising 242 groups. This enhancement in prediction accuracy is accompanied by a reduction in RMSE of over 50 % when compared to a solitary XGB model. A high R2 value indicates that the model possesses exceptional explanatory power and predictive accuracy, while the 95 % confidence interval provides reliable uncertainty boundaries for corrosion risk assessment and safety margin determination in engineering practice. The interpretability of the model was enhanced through the implementation of Shapley Additive Explanations (SHAP) and KAN weight analysis, which facilitated the visualization of both global and local feature contributions. The findings suggest that the water content (wc), dissolved chloride ions (cc), pH, and the interaction feature wc_rp exert a substantial influence on pipeline corrosion. This model achieves a balance between predictive accuracy, interpretability, and uncertainty quantification capabilities, thereby providing a reliable foundation for decision-making processes regarding pipeline corrosion monitoring and maintenance in scenarios involving small sample sizes. • Development of a hybrid physics–data-driven deep learning framework for small-sample pipeline corrosion depth prediction. • First application of Kolmogorov–Arnold Network (KAN) in corrosion prediction for high-order residual feature embedding. • Integrating BN-based causal feature selection with mechanism-driven interactions to improve physical consistency. • SHAP and KAN attention analyses highlight pH, chloride ions, and pipe-to-soil potential as primary corrosion drivers. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Loss Prevention in the Process Industries is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)*