A multiscale framework combining pore-level thermodynamics and process simulation via Bayesian-informed deep learning for CO2 adsorption

Accurate modeling of carbon dioxide (CO2) adsorption units under high-pressure conditions requires the integration of mass, energy, and momentum balances with a rigorous thermodynamic description of adsorption equilibrium. In this context, this work proposes a multiscale framework in which pore-level thermodynamics are provided by an extended Peng–Robinson equation of state for confined fluids (PR-C), supported by high-pressure adsorption data experimentally obtained using a Magnetic Suspension Balance (MSB). The PR-C confinement parameters are estimated through Bayesian inference, enabling the quantification of uncertainty and the generation of statistically significant adsorption equilibrium datasets. These datasets are then used to train a Deep Neural Network (DNN) surrogate model that preserves the physical behavior captured by the PR-C formulation while introducing uncertainty quantification through dropout-based neural networks within macroscopic fixed-bed simulations. The proposed framework achieved good agreement with experimental isotherms and benchmark breakthrough data, reaching a coefficient of determination (R2) of 0.9986 across cross-validation folds. Simulation results highlight the influence of operating conditions, such as pressure, temperature, and feed composition on adsorption equilibrium and system dynamics. Overall, the methodology offers a robust and efficient platform for adsorption process modeling, with direct applicability to process optimization, risk assessment, and the development of CO2 capture technologies.