A supervised machine learning approach with dual GAN augmentation and Bayesian HPO for corrosion prediction of Mg-Li alloys

The current study overcomes the limitations of Mg-Li alloys by creating a dataset with 450 data points sourced from literature. Two generative adversarial network (GAN) architectures, non-saturating GAN (NSGAN) with residual generator blocks, spectral normalization and feature matching loss, and Wasserstein GAN with gradient penalty (WGAN-GP) were used to generate augmented datasets. WGAN-GP had relatively better attribute fidelity to the KS/correlation values with composite score of 0.641 than NSGAN augmented data with score of 0.629. NSGAN-augmented data provided ML R 2 mean of 0.752 than WGAN-GP with mean of 0.716 across all corrosion targets such as E corr , i corr , CR and H 2 evolution rate. The NSGAN augmented dataset were combined with experimental data and Bayesian hyperparameter optimization was performed using Optuna m-TPE across five ML models, namely RF, XGBoost, SVR, AdaBoost and CatBoost. The integrated Experimental+NSGAN+CatBoost-HPO framework obtained test R 2 values of 0.867, 0.814, 0.814 and 0.763 for E corr , i corr , CR and H 2 evolution rate, respectively, which are 12.5%, 13.7%, 13.8% and 15.3% higher than experimental CatBoost-HPO baseline. SHAP analysis and permutation importance identified crystal structure as the dominant predictor for all targets. This framework establishes NSGAN-based augmentation as a preferred approach for electrochemical data of Mg-Li alloys.