Abstract:

Ammonia energy catalytic combustion technology has become a hot topic in clean energy research due to its high thermal efficiency and low pollution characteristics. The core challenge lies in the development of efficient catalysts. However, traditional catalyst design predominantly relies on experience-driven trial-and-error methodologies, which are often plagued by extended development cycles and elevated costs. A machine learning (ML)-enabled approach aimed at the targeted development of catalysts has been introduced. By integrating both experimental data and literature findings, two comprehensive databases were constructed: one comprising 597 samples related to ammonia conversion rates and another containing 529 samples pertaining to nitrogen selectivity. Both databases encompass various catalyst compositions (e.g., Fe₂O₃, CuO, CeO₂) as well as reaction parameters (e.g., temperature, equivalent ratio). Three ensemble learning models,random forest regression (RFR), gradient boosting decision tree (GBDT), and categorical boosting (CatBoost),were employed to predict catalytic performance. The results demonstrated that the RFR model exhibited superior comprehensive performance in dual-target prediction. On the test set, the model achieved a coefficient of determination (R²) of 0.912 with a mean absolute error (MAE) of 0.047 for ammonia conversion rate prediction, while for nitrogen selectivity prediction, it attained an R² of 0.918 and MAE of 0.033, indicating enhanced accuracy with reduced prediction errors. Feature importance analysis revealed reaction temperature as the dominant factor influencing both ammonia conversion rate and nitrogen selectivity. Partial dependence plot (PDP) analysis uncovered significant synergistic effects between CuO and CeO₂ loadings. The model predicted nonlinear correlations between bimetallic loading ratios and catalytic performance at 500, 700, and 900℃. Experimental validation confirmed strong agreement between predicted and measured values, with prediction errors below 3%, achieving precise catalyst performance evaluation. This machine learning-driven paradigm revolutionizes traditional catalyst development approaches, reducing development cycles by over 60% and substantially lowering costs. The established transferable methodological framework provides critical insights for designing catalytic systems in clean energy applications, demonstrating significant engineering value through its generalizability and operational efficiency enhancement.

Key words: machine learning, catalyst, oxidation, performance prediction, optimal design

摘要：

氨基能源催化燃烧技术因高热效率与低污染特性成为清洁能源研究热点，其核心挑战在于高效催化剂的开发。针对传统试错法周期长、成本高的问题，提出数据驱动策略，整合597组氨气转化率与529组氮气选择性数据，采用随机森林回归（RFR）、梯度提升决策树（GBDT）及类别型梯度提升（CatBoost）模型进行性能预测。结果表明，RFR模型在双目标预测中综合表现最优，测试集对氨气转化率预测R²=0.912，MAE=0.047，氮气选择性R²=0.918，MAE=0.033。特征重要性分析显示，反应温度对两目标特征影响最大。通过部分依赖图（PDP）解析CuO与CeO₂的协同效应，预测在500、700及900℃下不同比例负载量对应的催化氧化性能，实验验证显示预测误差<3%。该方法显著缩短了开发周期、降低了成本，可为机器学习驱动的催化剂设计提供可借鉴的方法论框架。

关键词: 机器学习, 催化剂, 氧化, 性能预测, 优化设计