Explainable machine learning prediction of corrosion inhibition efficiency from molecular descriptors using virtual sample generation

• M. Akrom¹, S. Rustad¹, H.K. Dipojono², R.P.P. Sukanli³ and K. Hongo³

¹ Research Group for Quantum Computing and Materials Informatics, Faculty of Computer Science, Universitas Dian Nuswantoro, Semarang 50131, Indonesia
² Quantum and Nano Technologies Research Group, Institut Teknologi Bandung, Bandung 40132, Indonesia
³ Research Center for Advanced Computing Infrastructure, JAIST, Nomi, Ishikawa, 923-1292, Japan

Abstract: The discovery of effective corrosion inhibitors is often constrained by the availability of experimental data, which limits the development of reliable predictive models. In this study, a machine learning framework is developed to predict corrosion inhibition efficiency (IE) using molecular descriptors derived from the Simplified Molecular Input Line Entry System (SMILES) representations. Unlike conventional approaches that often employ feature reduction, this study retains all 208 molecular descriptors and mitigates the limitations of small datasets by using mixup augmentation to simulate virtual samples. The results show that increasing the number of virtual samples consistently improves the predictive performance of ensemble models, including Random Forest (RF), Bagging, and Gradient Boosting (GB), with the coefficient of determination (R²) increasing significantly and steadily as the dataset size increases. To improve model interpretation, explainable artificial intelligence (XAI) is applied using SHAP analysis. The results indicate that molecular topological descriptors, such as Chi4n, Chi3v, Chi1n, and Chi4v, are the most influential features, while MolMR, which is related to electronic polarisation, also contributes significantly to predicting inhibition efficiency. These findings are consistent with a corrosion inhibition mechanism involving molecular topology and electron distribution that influence adsorption behaviour on metal surfaces. Additional analyses, including correlation analysis, partial dependence plots, and Williams plot evaluation, confirmed that the developed model provides reliable predictions in valid application domains. Overall, the proposed framework demonstrates the potential of integrating virtual sample simulation, machine learning, and model interpretation to accelerate the discovery and design of corrosion inhibitors within a materials informatics paradigm.

Keywords: organic corrosion inhibitor, machine learning, mixup data augmentation, explainable artificial intelligence, molecular descriptors

Int. J. Corros. Scale Inhib., 2026, 15, no. 3, 473-501
doi: 10.17675/2305-6894-2026-15-3-21

Download PDF (Total downloads: 28)

Back to this issue content: 2026, Vol. 15, Issue 3 (pp. 1-... (in progress))