Neural network‐assisted multi‐objective optimization of dual‐reflux pressure swing adsorption in biogas upgrading

Abstract BACKGROUND Dual‐reflux pressure swing adsorption (DR‐PSA) is an advanced variant of conventional PSA systems, capable of simultaneously producing two high‐purity products with high recovery. However, inherent trade‐offs among purity, productivity, and cycle work necessitate multi‐objective optimization (MOO) to balance performance. This study investigates the MOO of the DR‐PSA process for biogas upgrading using artificial neural networks (ANNs) as surrogate models. Single‐hidden‐layer ANNs were developed with various structures and training algorithms and integrated into the non‐dominated sorting genetic algorithm‐II to solve MOO problems. RESULTS Different DR‐PSA performance metrics require distinct ANN architectures and algorithms, indicating no universal model selection. The Adam‐ selu‐ 446 model for CH 4 purity in light product (LP) and the Adam‐ relu ‐398 model for cycle work, both developed on a large dataset, exhibited high accuracy, with of 0.9975 and 0.9959, respectively. Meanwhile, the Adam‐ elu ‐212 model for bed productivity, developed using a small dataset, achieved an of 0.9988. The Pareto front for maximizing CH 4 purity in LP and bed productivity revealed a sharp knee region, indicating a significant trade‐off. Under constrained bed productivity and cycle work, distinct Pareto fronts emerged at different CH 4 purity levels (>91% and 91.5%), with fewer feasible solutions at higher purity. Ultimately, the optimized conditions resulted in an enhancement in CH 4 purity in LP by at least 17%, a reduction in bed productivity by as much as 34%, and a reduction in cycle work by at least 30%, compared to the base case. CONCLUSION ANN‐assisted MOO optimizes DR‐PSA performance, offering insights into ANN model development, Pareto front mapping, and identification of feasible operational regions. © 2025 Society of Chemical Industry (SCI).