Skip to content
Review

Machine learning-assisted prediction and SHAP-based analysis of non-metal-doped ZnO photocatalysts: Data-driven modelling of bandgap, degradation efficiency, and kinetics.

Nov 2026 · Journal of Molecular Graphics and Modelling · Vol 150, pp. 109649 · 0 citations · 31 references
Medicine

Abstract

The doping of non-metals such as N, C, S, F, B, P, Cl, I, Si, and Se into ZnO is a well-established method to reduce the intrinsic wide bandgap of ZnO and broaden its optical activity into the visible region, which is beneficial for its photocatalytic degradation of organic pollutants. A collection of experimentally reported non-metal doped ZnO systems (400 in total) from peer-reviewed literature was selected to train machine learning (ML) models that predict three technologically critical properties: optical bandgap (Eg), photocatalytic degradation efficiency (%), and pseudo-first-order rate constant (k). A set of twelve compositional, structural, and process-condition descriptors were used for training and evaluation of eight regression algorithms: Linear Regression, k-Nearest Neighbors, Support Vector Regression, Artificial Neural Network, Random Forest, Gradient Boosting, XGBoost, and LightGBM, via 5-fold cross validation. Gradient Boosting and XGBoost consistently gave the highest cross-validated R2 under the original protocol, although the margin over tuned target-scaled SVR and ANN is small, the coefficient of determination (R2) being 0.990, 0.986, and 0.982 for Eg, degradation efficiency and rate constant, respectively. SHAP (SHapley Additive exPlanations) analysis indicated that the bandgap model relies mainly on surface area, dopant identity, and dopant concentration, whereas the degradation-efficiency and rate-constant models rely mainly on reaction time, followed by initial dye concentration. These rankings describe which inputs the trained models use and are not proof of physical mechanisms; in particular, reaction time acts mainly as an indicator of the experimental conditions (a model without it retains most of its accuracy), and material descriptors alone have almost no predictive power for degradation efficiency and k. The reported R2 values come from random 5-fold cross-validation of literature-compiled data, without DFT calculations or independent experimental validation, and should therefore be read as in-distribution accuracy. This work presents a data-driven correlation framework (three separately trained single-target models with a common workflow) for three performance indicators across ten non-metal dopants, intended to support hypothesis generation and preliminary prioritisation among the ten studied dopants for subsequent first-principles and experimental study; because the models use post-synthesis specific surface area as an input and do not extrapolate to unseen dopants, they are not pre-synthesis predictors. Leave-one-dopant-out validation gave a negative R2 for Eg, i.e. the models do not extrapolate to unseen dopants.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.