Prediction of corrosion weight loss in CoCrFeNiMo high-entropy alloys using multiple machine learning models (RIP2026-00006)

Roman Bogdanov

Machine learning (ML) are widely used to predict the corrosion resistance of various types of alloys including so-called high-entropy alloys (HEAs).Although some authors use complex ML models like neural networks, they often rely on short-term tests and qualitative parameters.This approach does not allow for the quantitative determination of mass loss and, consequently, corrosion rates, limiting long-term predictions.This study employs ML to predict the long-term corrosion of CoCrFeNiMo HEAs.Using gravimetric test data,various ML models were trained and optimized via 5-fold cross-validation and grid search. XGBoost demonstrated the highest accuracy on the test set, with the lowest MAE, while the Decision Tree (DT),being simpler,showed less overfitting.The models split data based on molybdenum content and exposure time, with XGBoost using complex trees involving other elements.All models tended to underestimate weight loss values.Validation on an additional dataset revealed distinct prediction patterns: XGBoost showed gradual weight loss changes with high variability,Random Forest (RF) predicted sharp rises followed by stabilization,and DT provided stable predictions of the weight loss values.Mo and exposure time were key features for all models. For RF, Fe also significantly influenced predictions. Further validation with experimental data will improve model accuracy and expand the training database, leading to more reliable corrosion-resistant alloy designs.