Adsorption processes see widespread use in industrial gas separations, including gas purification and CO2 capture [1]. Accurate modelling of these processes requires solving tightly coupled nonlinear partial differential equations (PDEs) that describe the mass, momentum, and energy transport. These PDEs are hyperbolic and highly stiff due to nonlinear adsorption equilibria. Moreover, the process is inherently transient and modular in nature and operates at cyclic steady state (CSS). Consequently, process simulation entails repeatedly solving PDEs over multiple iterations, from an arbitrary initial state until CSS is reached. Traditional simulators therefore rely on advanced numerical methods for reliably predicting process performance, resulting in high computational cost. Hence, there is a need for faster solution methods. Physics-informed neural networks (PINNs) have emerged as a promising approach for learning and accelerating PDE-based systems [2, 3]. Recent studies have demonstrated their capability for predicting the dynamics of vacuum-swing adsorption CO2 capture process [1]. However, training PINNs for adsorption processes remain challenging for the following reasons: Strong nonlinearity, tight coupling between the PDEs, and differing variable scales impede PINN learning of the coupled dynamics. Soft-constrained PINN formulation introduces multiple loss terms, making loss-weight tuning and convergence non-trivial. Finally, sharp concentration fronts and shock-like features further exacerbate the learning difficulties. We first developed a PINN model to simultaneously learn mass and energy transport in an adsorption column, with results shown in Fig. 1. However, successful training to a satisfactory accuracy proved to be a great challenge, for the above reasons. Recently, several strategies were proposed to address these limitations [4], including physics-based reformulations, training-level modifications, and architectural enhancements. Methods such as Stiff-PINN [5] and Characteristics-Informed Neural Networks [6] are of relevance. In this study, we systematically investigate and quantify the impact of PINN architectural modifications and physics-based reformulations on training efficiency and predictive accuracy for adsorption processes.