Deep Learning Surrogate Models for High-Pressure CO2 Adsorption: A Bayesian Approach with Extended Peng-Robinson Equation

Deep Learning Surrogate Models for High-Pressure CO2 Adsorption: A Bayesian Approach with Extended Peng-Robinson Equation

• Autor
• Marlon de Souza Gama
• Co-autores
• Felipe Rocha Pinto , Carla L. Manske Camargo , Frederico Wanderley Tavares , Amaro Gomes Barreto Jr.
• Resumo

• Accurate modeling for the removal of carbon dioxide (CO2) by adsorption entails the integration of principles of mass,

  energy, and momentum conservation with adsorption models. In this context, the thermodynamic description of

  adsorption equilibrium can be approached at various levels of detail. However, incorporating highly detailed isotherms

  presents a significant challenge for adsorption column solvers, as it requires solutions at each time step and position,

  leading to a substantial increase in computation time. To address this challenge, we developed a Deep Neural Network

  to train a surrogate model using adsorption data obtained through the solution of an extended Peng-Robinson model for

  confined fluids. These equilibrium data for CO2 adsorption were obtained by a Magnetic Suspension Balance (MSB),

  which enables sorption measurements at elevated pressures. We evaluated the extended Peng-Robinson model within a

  Bayesian framework to generate a collection of pseudo-experimental points, aiming to be used as inputs to train a model

  based on pattern identification techniques. The Bayesian structure is advantageous as it incorporates information from

  experimental data and prior assumptions of general knowledge (a priori knowledge) to generate posterior distributions

  of parameters and models. Finally, the output from the surrogate model is integrated into the mass balance equations

  while solving the partial differential equations that describe an adsorption column for CO2 removal. This approach

  results in a model for the evolution of the mass front in a fixed bed, facilitating enhanced accuracy in predicting system

  behavior.

• Palavras-chave
• Carbon dioxide removal, Multiscale model, Confined fluids, Deep Neural Network, Bayesian statistics
• Área Temática
• Modelagem e Simulação