Machine learning assisted evaluation of emission characteristics in oxy-fuel combustion for a supercritical CO2 power cycle

• CO emissions are analyzed in a 50 MW supercritical CO 2 combustor up to 17 MPa. • SVR with SHAP is used to interpret emissions in strongly coupled closed-cycle data. • CO emissions are dominated by outlet temperature; pressure effects are weak. • A two-zone finite-residence-time equilibrium model is developed from SHAP insights. • The apparent pressure–CO trend is attributed to covariation in high-pressure tests. This study analyzes exhaust emissions from a 50 MW direct-fired supercritical carbon dioxide (CO₂) combustor operated at pressures up to 17 MPa in a closed supercritical CO₂ power cycle. To disentangle tightly coupled operating parameters, equilibrium calculations, correlation analyses, and an explainable machine-learning model are employed. In the closed cycle, strong interdependencies among operating parameters make causal relationships difficult to identify from the raw data alone. A support vector regression (SVR) model is trained to predict carbon monoxide (CO) concentration and is interpreted using Shapley additive explanations (SHAP). The results show that combustor outlet temperature is the dominant factor governing CO formation, while pressure has a secondary influence. The trained model reveals a non-monotonic temperature effect: increasing temperature first decreases CO, and further increases beyond a threshold slightly increase CO. Based on these insights, the equilibrium model is improved by incorporating a two-zone, finite-residence-time effect. This improved equilibrium model reproduces the observed temperature dependence and shows better agreement with measurements. Reanalysis of the experimental dataset shows that the initially observed pressure–CO increase arises from covariation among operating conditions in the closed cycle rather than a direct pressure effect. These findings clarify emission characteristics of supercritical CO₂ oxy-fuel combustion and demonstrate that explainable machine learning helps avoid misinterpretation in strongly coupled systems.