Abstract
This work explores symbolic regression (SR) as an interpretable surrogate modelling approach for the multi-objective optimisation of pressure swing adsorption (PSA) systems for CO2 capture. A first-principle model was used as a virtual plant to generate synthetic datasets covering the operating space defined by cycle step durations. Two surrogate frameworks were developed and compared: SR models derived through evolutionary search and deep neural networks (DNNs) trained via Hyperband-based tuning. Both surrogates were used as simulation models within an optimisation procedure based on a particle swarm optimisation (PSO) algorithm to maximise CO2 purity and recovery. While DNNs achieved the lowest prediction errors (MSE ˜ 10-6), the SR surrogates provided compact analytical representations and significantly faster optimisation. The SR framework yielded a denser and more diverse Pareto front (4345 vs 508 points). It was about 34 times faster (38.6 s vs 1331 s), confirming its efficiency for surrogate-based optimisation of cyclic adsorption processes.