Experimental and AI-driven enhancements in gas-phase photocatalytic CO2 conversion over synthesized highly ordered anodic TiO2 nanotubes

• Published in: Energy conversion and management Vol. 327; p. 119544
• Main Authors: Hossen, Md. Arif, Hasan, Md. Munirul, Ahmed, Yunus, Aziz, Azrina Abd, Yaacof, Nurashikin, Leong, Kah Hon
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2025
• Subjects: administrative management; Artificial neural networks (ANN); automation; carbon dioxide; CO2 photoconversion; crystallites; energy conversion; fuels; hydrogenation; irradiation; light intensity; Machine learning (ML); nanotubes; neural networks; Photocatalysis; photocatalysts; prediction; surface area; Titania nanotube arrays (TNTAs); titanium dioxide; value added; CO2 photoconversion; Machine learning (ML); Titania nanotube arrays (TNTAs); Artificial neural networks (ANN); Photocatalysis
• ISSN: 0196-8904
• DOI: 10.1016/j.enconman.2025.119544

[Display omitted] •Six machine learning models were applied to predict gas-phase CO2 conversion efficiency over TNTAs photocatalysts.•The K-fold cross validation method tuned the hyperparameters.•The ANN model outperformed other models with R2 values of 0.983 and 0.984 for training and testing datasets.•The maximum CO2 photoconversion was achieved at 71.32 % under optimal experimental conditions.•The exposed surface area of catalyst was found to be the most influential photocatalytic input parameter. The photocatalytic hydrogenation of CO2 to value-added products is one of the most appealing sustainable strategies to meet growing fuel demand and lowering CO2 levels in the atmosphere. This study focuses on the prediction and optimization of CO2 conversion efficiency using machine learning (ML) approach over synthesized highly ordered TiO2 nanotube arrays (TNTAs) photocatalysts. Six popular ML algorithms of regression, kernel and neural network-based models were applied to predict the gas-phase CO2 photoconversion rate. The percentage of CO2 conversion was taken as the targeted feature while catalyst exposed surface area, crystallite size, CO2 concentration, light intensity, feed pressure, and irradiation time were chosen input parameters. The K-fold cross-validation method was employed to fine-tune the hyperparameters of the ML models. Compared to other ML models, the artificial neural networks (ANN) model showed an outstanding rating of performance evaluation metrics, including R2, MSE, RMSE, and MRE. Consistency of performance indicator values were also revealed during the testing of models. These results demonstrated the robust and effective predictive capabilities of the ANN model for CO2 conversion efficiency. The optimization of the input parameters for CO2 photoconversion was comprehensively validated using the predicted and experimental data. The highest prediction achieved for the photoconversion of CO2 was 71.32 %, while experimental validation confirmed an effectiveness of 70.60 % under optimized conditions. The SHapley Additive exPlanations (SHAP) method and normalized importance showed that the parameters related to photocatalysts were the most influential features in enhancing the gas-phase CO2 photoconversion compared to parameters related to photocatalytic settings. The integration of ML approach and gas-phase photocatalytic CO2 conversion has immense future potential for better design of other photocatalytic solar-based energy production applications. Future investigations should concentrate on the integration of ML predictions with real-time monitoring systems to facilitate process automation, adaptive adjustments, and improved scalability for practical applications.