Data-augmenting self-attention network for predicting photocatalytic degradation efficiency: a study on TiO2/curcumin nanocomposites

Abstract

In a previous work, a cellulose paper film containing a nanocomposite charge (TiO₂/5%-curcumin) was developed and used in the photocatalytic degradation of methylene blue (MB), an organic dye, on a tubular reactor with trickling and circular flow under UV irradiation light. The effect of three main operational parameters on the photocatalytic degradation of MB was studied: the mass of TiO₂/5% Curcumin material deposited on the cellulose paper, the initial concentration of the pollutant (MB) and the intensity of UV irradiation light. The obtained results show that by working under operating conditions of mass of deposited material (14 mg), initial pollutant concentration (10 ppm) and intensity of UV irradiation light (3.76 w/cm²), approximately 85% of MB was removed after 220 min of irradiation. To help guide future experimental efforts, the present work proposes a data augmenting self-attention network (DASAN) for the prediction of the photocatalytic degradation efficiency from a set of experimental parameters. The suggested ensemble combines base models for data augmentation and a meta model incorporating a self-attention mechanism for the prediction. The model is trained using the obtained experimental data of operating conditions (mass of material, initial pollutant concentration and intensity of UV irradiation light). The base models achieved excellent fits to the data and the meta model attained a mean squared error of 0.0055 through five-fold cross-validation, predicting optimal degradation efficiencies of 86–90% for experimental values of 20–22 mg the catalyst charge, 11–15 ppm for the initial pollutant concentration and 4.4–5.7 w/cm² for the intensity of UV irradiation light.