Global sensitivity analysis of blue hydrogen production: a comparative study using machine learning

Data-driven modelling utilising machine learning (ML) techniques offers a powerful alternative to first-principles simulations of chemical processes. In this work, artificial neural networks and random forests were developed as surrogate models, trained on data from a first-principles model of sorption-enhanced steam methane reforming with chemical-looping combustion. These ML-based surrogates were integrated with global sensitivity analysis (GSA) approaches to identify key process drivers and evaluate the comparative performance of different GSA methods in chemical process modelling. The surrogate models achieved an approximately 99 % reduction in computational time compared to first-principles simulations, while maintaining predictive accuracy. Sensitivity analysis demonstrated that the CaO/natural gas (CaO/NG) ratio is a dominant parameter, strongly influencing carbon capture efficiency and hydrogen production performance (cold-gas efficiency and H 2 purity). In-situ CO 2 removal from the reformer was shown to shift equilibrium towards higher hydrogen yields while simultaneously enabling CO 2 capture. Ratios of CaO/NG ≥ 1.00 ensured high capture efficiency, while improvements in cold-gas efficiency were observed from ratios ≥0.5. Among GSA methods, the Sobol approach delivered high computational efficiency (0.5 s) with first- and second-order sensitivities, whereas Shapley additive explanations provided greater interpretability but at significantly higher computational cost (384 s). • Large dataset (32,628 points) used to train a surrogate model for blue hydrogen production. • Machine learning surrogate modelling reduced computational time by about 99 %. • Global sensitivity analysis shows higher CaO/natural gas ratio improves H 2 yield and CO 2 capture. • Shapley additive explanation shows limited CO 2 capture gains beyond CaO/natural gas = 1.0.