Membrane gas separation is a technology used in highly relevant industrial applications, such as hydrogen recovery from ammonia plants and refineries or CO₂ capture from power plants and industrial emissions. In these processes, materials with specific gas permeability are exploited to separate different components.

The performance of this technology depends on various factors, such as material properties, membrane areas, pressure, etc. Moreover, several separation stages might be needed to meet the requested levels of recovery and purity. The goal of optimal process design is to determine the configurations of minimum separation cost. In this work, this problem is modeled as a global optimization problem, in which the single-stage membrane behavior is represented by an artificial neural network (ANN). A similar problem was already studied (Knitro – Artelys) with a different modeling strategy. In this new work, the authors propose an efficient data augmentation method to improve the ANN prediction based on simulations around local optimal solutions.

The authors rely on Artelys Knitro to solve the nonlinear optimization problems (NLP) arising both in the global optimization strategy (multistart) and in the data augmentation procedure. This shows how Artelys Knitro can be effectively employed to design these processes and to guide the data augmentation algorithm.