A study on the prediction of normalized concentration in photocatalytic degradation of 2,4‐dinitrophenol using multilayer perceptron models

Abstract 2,4‐dinitrophenol (2,4‐DNP) is commonly found in industrial effluents and poses serious risks to humans and the environment, prompting various investigations into its treatment and development of reliable predictive models. In this study, machine learning regression models have been developed to analyze the photocatalytic degradation performance of 2,4‐DNP based on measured experimental data for the NH 2 ‐MIL‐125(Ti)/red phosphorus photocatalyst. Neural network models were developed to predict its normalized concentration ( C / C 0 ) with respect to the main operational conditions, including the initial concentration of the pollutant, catalyst dosage, the pH of the solution, and the reaction time. Multilayer perceptron architectures have been investigated with an exhaustive evaluation of 672 network architectures by varying network depth, distribution of neurons, activation function, and optimizer algorithm. The coefficient of determination ( R 2 ) and mean squared error (MSE) have been used to evaluate model performance, which is finally evaluated using independent test data. The optimal network configurations consistently employ the hyperbolic tangent activation function in combination with the L‐BFGS optimization algorithm, achieving high predictive accuracy ( R 2 > 0.99) with minimal prediction error (MSE ≈ 2.8 × 10 −5 ). The analysis of feature importance exhibits the solution pH and reaction time as the two most dominant factors for determining the degradation behaviour, whereas initial pollutant concentration and catalyst dosage have a relatively weaker effect on the degradation behaviour over the studied experimental conditions. In general, the findings indicate that the MLP model can be regarded as a useful surrogate in predicting and analyzing the photocatalytic degradation for advanced wastewater treatment processes.