Physics-Guided Temporal Transformer for CO????Corrosion with SHAP Case study: Asaluyeh SPGC Complex, Iran

The structural integrity of the oil and gas industry is at high risk due to CO₂-induced corrosion; consequently, practical, explainable mechanisms are required to accurately predict this phenomenon. The available mechanistic models, like the de Waard–Milliams equation, though entrapping fundamental thermodynamics, do not withstand dynamic dosing and time-varying conditions. Although the conventional machine learning and deep learning models improve accuracy, though often lack physics consistency and interpretability. A Physics-Guided Temporal Transformer (PGTT), which integrates a Temporal Fusion Transformer backbone with a physics-informed loss (of de Waard–Milliams), and SHapley Additive exPlanations (SHAP) for feature attribution, is proposed here. By applying a comprehensive sequential dataset from 22 inhibitor dosing experiments (15,400 samples during 60+ hours), this PGTT achieves higher performance with an average , , and of over five independent runs, and outperforms Random Forest , , and MLP baselines. According to the SHAP, Temperature (26.5%) and CO₂ pressure (17.7%) are identified as the dominant stimulants, consistent with corrosion science. This case study reveals a reduction in prediction error of less than 5%, supporting proactive inhibitor dosing and pipeline integrity management.