Development of Data-Driven Hybrid Models for Fixed-Bed Gas Adsorption Processes

Abstract

This dissertation investigates hybrid modelling strategies for fixed-bed gas adsorption processes, with emphasis on the use of data-driven methods combined with first-principles models. Fixed-bed adsorption is a relevant unit operation in gas separation and purification, and its dynamic behaviour is governed by coupled mass, energy, and momentum transport phenomena that are often difficult to represent with full accuracy using purely mechanistic formulations. A one-dimensional mechanistic model was developed and implemented in Python, and its predictions were validated against experimental data and gPROMS simulations. Building on this baseline, two hybrid approaches were explored. For equilibrium separations, a Symbolic Regression superposition model was used to correct systematic temperature prediction errors. The analysis showed that the temperature residual had a strong autoregressive structure, and the resulting correction model significantly improved agreement with the experimental thermal histories. For kinetic separations, a NeuralODE framework was developed by replacing the classical Linear Driving Force uptake-rate expression with a small neural network embedded in the adsorption model. After fitting the equilibrium behaviour with a Dual-Site Langmuir model, the NeuralODE improved breakthrough prediction for CH4/N2 separations, even when trained on limited experimental data. The results show that hybrid models can preserve the physical structure of adsorption simulations while improving predictive accuracy in both thermal and kinetic descriptions. The proposed frameworks provide a flexible basis for further development of interpretable and data-enhanced models for adsorption processes.