Predictive Modelling of Corrosion Inhibition from Eggshell-Derived Solutions Using Robust Machine Learning Algorithms

Corrosion inhibitors derived from agricultural waste, particularly eggshells, are gaining traction as an economically viable, green alternative to traditional inhibitors. Yet, measuring corrosion rate in experimental settings remains resource- and time-intensive and challenging, thereby warranting the application of data-driven predictive models. This paper presents an intensive, reliability- and interpretability-focused approach for rapidly predicting the performance of corrosion inhibitors using machine learning models. Literature records were curated with documented exclusion rules, including removal of a series whose corrosion rates contradicted its own reported inhibition efficiency, giving 151 observations across five alloy–medium systems. Ridge-regression and nearest-neighbour baselines and six learners (Decision Tree, Random Forest, Support Vector Regression, XGBoost, LightGBM and Multilayer Perceptron) were compared using a stratified 80/10/10 split, nested repeated cross-validation, bootstrap confidence intervals, grouped cross-validation, Williams-plot applicability-domain analysis and Shapley Additive Explanations. Under the pre-defined criterion of lowest nested cross-validation RMSE, Random Forest was selected (0.101 ± 0.064 mmpy; R² = 0.943 ± 0.044), with XGBoost statistically indistinguishable (0.107 ± 0.065 mmpy). On the internal test and held-out validation sets (n = 16 each), Random Forest gave R² = 0.991 (95% CI 0.838–0.997) and 0.931 (0.316–0.992); these wide intervals show why small hold-out sets cannot support broad generalisation claims. Grouped cross-validation confirmed reliable interpolation to unseen exposure times (R² = 0.96) and, less accurately, to unseen inhibitor doses (RMSE = 0.11 mmpy), but not to uninhibited behaviour or, reliably, to unseen alloy–medium systems. Alloy–medium identity and inhibitor concentration dominated the predictions, with corrosion rate decreasing with increasing dose. Finally, the results analysis correctly matched the experimental results: higher temperature reduced the effectiveness of the inhibitors, while higher inhibitor concentration increased their ability to resist corrosion due to increased surface coverage. The framework is an interpolative screening tool that complements, rather than replaces, kinetic, Arrhenius and adsorption-isotherm analysis.