Uncertainty-Aware Hydrogen Production Optimization from a Steam Methane Reformer with a Water Gas Shift Reactor and Pressure Swing Adsorption

This paper presents a novel framework for integrated hydrogen production using machine learning models with uncertainty quantification. The proposed method is capable of handling the variation for the hydrogen process chain to accurately predict the net annual hydrogen production with uncertainty estimates. Detailed numerical models are developed for each process step, capturing the necessary mass, energy, and momentum balances. These high-fidelity simulations are validated against literature data and industrial measurements, demonstrating close conversion, recovery, and productivity predictions. To alleviate the computational overhead inherent in the numerical simulations, we train two machine learning (ML) surrogates–a multilayer perceptron (MLP) and an attention-based model (ATTN)–using the generated simulation data. Both surrogates incorporate uncertainty quantification to provide point predictions and confidence intervals, thus enabling risk-aware decision-making. Evaluation results indicate that the ATTN model consistently achieves higher accuracy and better-calibrated coverage than the MLP, with a coefficient of determination R² typically exceeding 0.98 and a prediction interval coverage probability (PICP) within 5% of the nominal coverage across multiple operating conditions. Furthermore, a sensitivity analysis conducted with the ATTN model reveals the critical influence of temperature, pressure, and feed composition on the production process and provides actionable insights for process design and optimization. These findings suggest that an attention-based surrogate, augmented with robust uncertainty estimates, can be a powerful tool for rapid decision-making and efficient hydrogen production optimization.