Modelagem e previsão da adsorção do CO2 em diferentes adsorventes utilizando equações fenomenológicas e redes neurais: uma abordagem combinada

Abstract

The high emission of carbon dioxide (CO2) is one of the main causes of global warming. Adsorption is one of the potential options for capture, since it is easy to apply and adapt in industrial processes, in addition to being low cost. However, it is still necessary to study the optimal conditions to increase capture efficiency. Among the materials used for adsorption, zeolites, metal-organic frameworks and activated carbon stand out, since their characteristics such as pore volume and distribution, as well as surface area are variables of significant importance for the efficiency of the process. However, in classical equilibrium modeling these properties do not appear in the mathematical formulation. Phenomenological models based on equations of state are alternatives to describe adsorption phenomena, however, the phenomenological route presents some limitations when the number of variables involved in the process increases. Thus, the main objective of this work is to perform the mathematical modeling of equilibrium data of CO2 adsorption on different types of adsorbents using phenomenological models (2D Equations of State and classical Adsorption Isotherms) and artificial neural networks. Thus, adsorption data collection was initially performed for the following adsorbents: Cu-BTC, Zeolite 13X, IRMOF-1, ZIF-8, Mg-MOF-74, activated carbon and Zeolite 5A, with different conditions of temperature, pressure, besides the textural properties of the adsorbents: surface area and pore volume, thus totaling 2991 data of temperature, pressure and adsorbed amount. After data collection, mathematical modeling was performed with the classical equations (Langmuir, Freundlich, Toth and Sips) and with the two-dimensional equations of state (Van der Waals, Redlich-Kwong and Pegn-Robinson). The models that best represented the investigated systems were Toth (IRMOF-1 and activated carbon), Sips (ZIF-8 and Zeolite 5A), Van der Waals (Zeolite 13X and Mg-MOF-74) and Langmuir (Cu-BTC). In a second step the best model for each type of adsorbent was used to generate a standardized database with the adsorption equilibrium conditions together with the textural properties to perform the training with neural networks. For the training, the K-Fold cross-validation technique was used, with 4 subsets, with 15% separation of the data to perform the final validation, 6 different conditions were tested and optimized, with 1, 2 and 3 internal layers of neurons, testing perceptron and recurrent neurons. The criterion used to separate the data was the combination of the input variables: area, temperature and pore volume for Cu-BTC, Zeolite 13X and IRMOF-1, and area and temperature for the rest of the adsorbents, in order to avoid statistical analysis with previously trained data and bring greater robustness in the models. The number of neurons and the activation function were chosen using genetic algorithms. With the best configuration chosen, the optimal number of epochs was then determined, comparing with the test data. Finally, with the best configuration, a statistical comparison was made to choose the best model obtained among the 6 configurations. In general, the configuration with perceptron neurons stood out in relation to the recurrent networks, and only for activated carbon that the modeling was not satisfactory. Through the parametric analysis it was observed that the area has a negative correlation with adsorption, except for Zeolite 13X and the pore volume showed a positive correlation for Cu-BTC, negative for Zeolite 13X and inconclusive for IRMOF-1. Furthermore, the use of neural networks combined with phenomenological equations was satisfactory for generating generic models with predictive capacity under different operating conditions and textural properties of the adsorbents.