Optimization on cleaner intensification of ozone production using Artificial Neural Network and Response Surface Methodology: Parametric and comparative study

Abstract

Abstract Non-thermal microplasma is a promising technology for efficient ozone generation for water sterilization. However, there remains a niche to improve the energy efficiency of the production process. The studies investigating the combined effects of interacting parameters affecting ozone generation are scarce. Studying more than one parameter is a limitation using standard experimental protocols. However, modeling tools such as Response Surface Methodology and Artificial Neural Network provides an opportunity to study the interaction between parameters and propose a mathematical model to predict ozone concentration under various experimental conditions. A robust model providing an insight into parametric interaction and better forecasting can reduce the required power requirement making it cleaner and sustainable. In this study, a Dielectric Barrier Discharge-Corona hybrid plasma microreactor, combining the homogeneity of the former and high energy streamers of the latter, was used to investigate factors affecting ozone generation. Response Surface Methodology was used with Central Composite Design for experimental design. A model was developed for analyzing the correlation of parameters, evaluate complex interactions among independent factors and optimization. Artificial Neural Network model based on Feed-Forward Backpropagation Network was developed to predict the response. The results were compared with the mathematical models developed by Response Surface Methodology. To the best of our knowledge, a study on ozone generation and optimization in a Dielectric Barrier Discharge-Corona hybrid discharge reactor do not exist. Similarly, such a detailed analysis and comparison of Response Surface Methodology and Artificial Neural Network for ozone generation is reported for the first time. Ozone generation was favored at lower values of flow rate and pressure of air, frequency, and higher voltage and electrode length. Response Surface Methodology was found to have a lower value of R2 = 0.9348 as compared to Artificial Neural Network, i.e., R2 = 0.9965. Root Mean Square Error obtained from Response Surface Methodology (5.0737) is approximately four times higher as compared to Artificial Neural Network (1.1779). The results showed the Artificial Neural Network model is more reliable than the Response Surface Methodology to study the interacting parameters and prediction. The model could be related to the real-time experiments to predict the ozone concentration under various experimental conditions and make the sterilization process cleaner.