A novel PINN-STAN model for predicting corrosion rate in subsea oil and gas pipelines

This paper presents a novel Physics-Informed Neural Network with Spatio-Temporal Attention Network (PINN-STAN) for predicting corrosion rates in subsea oil and gas pipelines. Traditional data-driven models fail to capture multi-factor coupling effects and lack physical consistency under extreme conditions. PINN-STAN overcomes these limitations by embedding corrosion kinetics and mass conservation laws as trainable constraints. It employs a Bayesian spatio-temporal attention mechanism to dynamically adjust the importance of environmental factors and uses a 1D Convolutional Neural Network (CNN) for short-term events and Neural ODEs for long-term trends. Additionally, the model integrates a partial differential equation (PDE) to describe the physical evolution of corrosion processes, ensuring physical consistency by coupling the PDE solution with deep learning features. Experimental results on the CNOOC dataset demonstrate that PINN-STAN achieves a Root Mean Squared Error (RMSE) of 0.0165mm/a and a coefficient of determination (R 2 ) of 0.9950, showcasing its superior predictive accuracy and stability under extreme conditions. Compared to traditional linear regression and machine learning models, PINN-STAN better incorporates physical constraints and multi-scale prediction capabilities, making it well-suited for corrosion assessment in multi-physical field coupling environments. This framework provides a reliable solution for corrosion prediction in complex subsea environments. • Developed a hybrid PINN-STAN framework for subsea pipeline corrosion prediction. • Integrated diffusion–advection–reaction PDEs to embed corrosion kinetics and physics. • Designed a spatio-temporal attention network to capture dynamic multi-factor coupling. • Demonstrated high robustness and generalization under complex marine conditions.