Ethylenethiourea presents substantial hazards to aquatic ecosystems and human health as a toxic, persistent degradation by-product of the widely applied agricultural fungicide mancozeb. However, establishing optimal reaction conditions using conventional empirical techniques poses significant challenges. Traditional trial-and-error optimization methods consistently struggle to capture the complex, non-linear interactions that govern operating variables in photocatalytic degradation systems.
To address these limitations, researchers developed a systematic modeling approach to evaluate the influence of key operating parameters on the photocatalytic degradation of ethylenethiourea. The investigation generated an empirical database comprising 448 distinct measurement datasets collected from 64 independent photocatalytic experiments. The study systematically trained, tested, and compared eleven distinct machine learning model architectures to assess their capacity to accurately predict ethylenethiourea removal efficiencies under diverse operational conditions.
Comparative performance evaluations revealed that an artificial neural network emerged as the superior predictive architecture among the eleven evaluated models. Evaluated on the independent test dataset, the artificial neural network achieved a coefficient of determination of 0.921, demonstrating robust correlation with observed degradation measurements. Additionally, the model recorded a mean absolute error of 2.84 alongside a root mean square error of 4.28, establishing high predictive precision and reliability across varied operational settings.
To overcome the black-box nature of machine learning algorithms, the investigators conducted interpretability evaluations using feature importance analysis and partial dependence plots on the optimal artificial neural network. The feature importance hierarchy established that reaction time served as the predominant governing parameter, accounting for 67.37 percent of total feature importance. Initial ethylenethiourea concentration ranked as the second most impactful operational variable at 21.15 percent, followed by reaction temperature at 8.43 percent, while titanium dioxide catalyst dosage contributed 3.05 percent.
Partial dependence plots provided detailed insight into the specific functional relationships connecting individual operating variables to degradation efficiency. The model revealed that increasing the reaction duration steadily elevated ethylenethiourea removal rates, whereas higher initial pollutant concentrations reduced overall removal efficiency. The influence of temperature displayed non-monotonic behavior: when the temperature increased from 30 °C to approximately 40 °C, partial dependence values declined to their minimum point, before reversing into an upward trajectory as temperatures exceeded 40 °C. Similarly, altering catalyst dosage produced a trend characterized by an initial rise in dependency values followed by a subsequent decrease.
Leveraging these captured relationships, the artificial neural network predicted an optimal combination of operating conditions designed to maximize ethylenethiourea removal. The model-derived parameters specified a total reaction time of 330 minutes, which incorporated 30 minutes of initial dark adsorption, an initial ethylenethiourea concentration of 7.5 milligrams per liter, a titanium dioxide catalyst dosage of 58.4 milligrams, and a controlled operating temperature of 30 °C. Under these specified conditions, the neural network projected a theoretical pollutant degradation rate of 75.31 percent.
The researchers subsequently verified the model’s predictive precision through five independent parallel laboratory validation experiments conducted under the optimal parameters. The physical experiments yielded an average ethylenethiourea degradation efficiency of 76.86 percent, showing close agreement with the model projection. The experimental validation demonstrated a sample standard deviation of 4.76 percent and a relative standard deviation of 6.19 percent, illustrating high experimental reproducibility and confirming that interpretable machine learning enhances predictive reliability in photocatalytic process optimization.
Subject of Research: Machine learning optimization of photocatalytic ethylenethiourea degradation by titanium dioxide
Article Title: Using Machine Learning to Investigate the Influence of Operating Parameters on the Photocatalytic Removal of Ethylenethiourea by TiO2: Feature Importance and Partial Dependence Analysis
Article References: Gu, H., Chang, J., Wei, S., Zhao, J., & Luo, Q. (2026). Using Machine Learning to Investigate the Influence of Operating Parameters on the Photocatalytic Removal of Ethylenethiourea by TiO2: Feature Importance and Partial Dependence Analysis. Catalysis Letters, 156(12), Article 335. https://doi.org/10.1007/s10562-026-05569-0
Image Credits: AI Generated
DOI: 10.1007/s10562-026-05569-0
Keywords: Ethylenethiourea, Photocatalytic degradation, Titanium dioxide, Artificial neural network, Machine learning, Mancozeb by-product, Feature importance, Partial dependence plots, Water treatment, Reaction time, Catalyst dosage, Pollutant removal, Process optimization, Catalysis Letters, Nonlinear modeling, Experimental validation
News Source: Blake Davidson. (November 19, 2026). Neural Network Model Optimizes Ethylenethiourea Photocatalytic Removal Using Titanium Dioxide. Scienmag.



