From Discharge to Data: Explainable Machine Learning Accelerates Green Nitrogen Fixation

Plasma-assisted nitrogen fixation offers a promising pathway for sustainable ammonia and nitrate synthesis under ambient conditions, yet its energy efficiency remains a major bottleneck. This study presents an interpretable machine learning framework that integrates ensemble learning and SHAP-based analysis to accurately predict and optimize the energy consumption of plasma–water-based nitrogen fixation (PWNF) systems. A total of 224 experimental observations from 21 discharge configurations were used to train and evaluate eight regression algorithms, including XGBoost, CatBoost, and SVR. The best-performing stacking ensemble, led by XGBoost as a meta-learner, achieved a test set R 2 of 0.966 and a RMSE of 10.17 MJ/mol, demonstrating excellent accuracy and generalization. SHAP and partial dependence plot analyses identified key variables─such as power, nitrogen percentage, discharge frequency, and water vapor content─and revealed their nonlinear contributions to energy efficiency. Furthermore, bivariate PDPs uncovered synergistic interactions (e.g., between power and gas composition) that guide multiparameter optimization strategies. Experimental validation confirmed the model’s robustness, with prediction errors of less than 7% for boundary cases. This framework not only advances predictive modeling in plasma engineering but also provides actionable insights into the parametric tuning of nitrogen fixation systems. The approach can be scaled to other plasma-assisted or hybrid electrochemical nitrogen conversion routes.