Hybrid pathways for methane production: Merging thermodynamic insights with machine learning

A comprehensive study was conducted to simultaneously simulate thermodynamic behavior and predict catalyst performance for CH₄ production via CO and CO₂ methanation, using blast furnace gas (BFG) and basic oxygen furnace gas (BOFG) as feedstocks. Thermodynamic equilibrium simulations based on Gibbs free energy minimization identified optimal reaction conditions at moderate temperatures (150–250 °C) and elevated pressures, achieving over 98 % CO₂ conversion with less than 1 wt% carbon formation. In parallel, machine learning models were developed using an augmented dataset of 2777 experimental observations. Atomic-level structural and electronic descriptors were incorporated into the dataset, including unit cell density and formation energy for active metals, promoters, and supports. Feature selection through Pearson correlation and RFECV identified active phase weight, support density, and reduction conditions as the most influential variables. Among all tested algorithms, XGBoost and CatBoost demonstrated the highest accuracy, with R² values exceeding 0.93 for predicting CH₄ yield, selectivity, and CO₂ conversion. SHAP and partial dependence analyses showed that catalyst stability and textural properties govern overall performance. This integrated thermodynamic and machine learning approach defines the operating limits for high-efficiency methanation and provides a data-driven framework for catalyst optimization in industrial applications.