The use of chemometric techniques and multivariate experimental designs for the photocatalytic reaction of reactive scarlet BES in aqueous solution under ultraviolet light irradiation is described. The efficiency of photocatalytic degradation was evaluated by the analysis of the parameter of decoloration efficiency determined by UV absorption at 540 nm using a UV-Vis spectrophotometer in different conditions. Five factors, such as the amount of titanium oxide ([TiO2]), the concentrations of reactive scarlet BES (c(0)), irradiation time (t), the pH value (pH) and temperature (T), were studied. [TiO2]. c(0), t and pH selected on the basis of the results of variance analysis by Plackett-Burman design were used as independent variables. Training sets and test sets of back propagation neural network (BPNN) were formed by Box-Behnken design and uniform design U10 (10 x 5(2) x 2) respectively. The process of photocatalytic degradation of the target object was simulated by the BPNN model. The correlation coefficient (r) of the calculation results for training set and test set by BPNN is 0.996 4 and 0.963 6 respectively, and the mean relative errors between the predictive value and experimental value of decoloration efficiency are 6.14 and 7.76, respectively. The modeled BPNN was applied to analyze the influence of four factors on decoloration efficiency. The results showed that the initial conditions of c(0) being lower, pH 5.0 and appropriate amount of [TiO2] contribute to improving the decoloration efficiency of reactive scarlet BES. Under the condition of c(0) = 40 mg x L(-1), the optimized experimental condition of the system was obtained: [TiO2] = 1.20 g x L(-1) and pH 5.0. Under the optimized experimental condition, the experimental value of decoloration efficiency is 98.20% when irradiation time is 35 minutes and the predictive value of decoloration efficiency is 99.16% under the same condition. The relative error of decoloration efficiency between the predictive value and experimental value is only -0.96%. The experimental value is very close to the model predicted value.