Bayesian Optimization for techno-economic analysis of pressure swing adsorption processes

Pressure swing adsorption (PSA) can remove CO₂ from flue gases. The full potential of the technology can only be exploited if the optimal combination of adsorbent material and process is identified. This identification requires screening the large database of adsorbent materials by performing computationally intensive process optimizations. To ensure a suitable compromise between accuracy and computational lightness, the PSA process can be described by reduced-order models. However, those models might involve several discrete states making the objective function discontinuous and not continuously differentiable, challenging gradient-based methods. The selection of suitable optimization methods is therefore an open issue. This study compares three optimization algorithms, Bayesian optimization, NOMAD and KNITRO, for two cases and adsorbent materials.

For the tested cases and adsorbent materials, none of the three algorithms is clearly superior to the other. Bayesian optimization (BO) needs fewest function evaluations and computational time to converge and outperforms the other two for one case. However, BO is less reliable than NOMAD and KNITRO for the other case tested.