This research presents an interpretable machine-learning model for predicting the corrosion rates of casing and tubing materials exposed to aggressive downhole environments across an ultra-wide range of conditions. The model is trained and evaluated on a consolidated experimental dataset of 744 High-Temperature and High-Pressure (HTHP) autoclave tests spanning extreme operating windows (e.g., temperature 25-220°C, CO 2 partial pressure 0-28.39 MPa, H 2 S partial pressure 0-2000 kPa, chloride concentration 0-160,000 ppm) and multiple material classes, including carbon steels and corrosion-resistant alloys (Cr 0-17.0%) that are representative of casing and tubing materials used in downhole conditions, and fluid types. Fourteen candidate learning algorithms were hyperparameter-tuned using particle swarm optimization (PSO). Among them, the TabPFN algorithm achieved the best generalization on an independent test set. To ensure transparency, SHAP (SHapley Additive exPlanations) was applied to quantify feature contributions and interactions. The combined TabPFN+SHAP analysis reproduces and explains complex corrosion behaviors—including the beneficial effect of chromium, the roles of temperature, CO 2 and H 2 S partial pressures, bicarbonate concentration, oil fraction and fluid type—consistent with corrosion electrochemical corrosion theory and microstructural observations. The framework bridges predictive accuracy and mechanistic interpretability and offers a practical decision-support tool for integrity management of casing and tubing materials in aggressive downhole environments.
Xing et al. (Wed,) studied this question.