Evidence-Based Design and Optimisation of Titania Photocatalysts via Artificial Neural Network Analysis

Abstract

This study deals with the development of artificial neural networkfor optimal titania-based photocatalyst design using over 700 data cases from the literature. In spite of the variability in intrinsic error across laboratories and continents over a 20-year span, feed-forward ANNs relating catalyst preparation variables to photocatalyst properties displayed good predictive capabilities with correlation coefficients generally greater than 0.92. Calcination temperature and dopant concentration exhibited strong negative connection weights to the optophysical properties of the catalyst (surface area, crystallite size, band-gap energy and point of zero charge) while dopant oxidation number and ionic radius have positive connection weights although the optimal ANN ensemblefor each photocatalystproperty contains different numberof neuronsin the hiddenlayer. A global ANN connecting both catalyst preparation variables and reaction conditions as inputs optimised the relationship to photoactivity with a 16-neuron hidden layer with calcination temperature, dopant concentration, molecular weight of the organic substrate and photocatalyst loading having a negative effect on photoactivty while the most important positive influence was provided by the initial concentration of the organic pollutant and dopant ioinic radius. Due to the large spectrum of input variables accommodated by this ANN, it may beusedasa meaningfulguideinthedesignofnew photocatalystsforspecific applications. Thereliabilityof this optimal ANN architecture is demonstrated with test data which has excellent parity plot with the predicted values.