Drivers of regional surface ozone bias drivers in chemical reanalysis air quality revealed by explainable machine learning

This study employs an explainable machine learning (ML) framework to examine the regional dependencies of surface ozone biases and their underlying drivers in global chemical reanalysis. Surface ozone observations from the Tropospheric Ozone Assessment Report (TOAR) network and chemical reanalysis outputs from the multi-model multi-constituent chemical (MOMO-Chem) data assimilation (DA) system for the period 2005–2020 were utilized for ML training. A regression-tree-based randomized ensemble ML approach successfully reproduced the spatiotemporal patterns of ozone bias in the chemical reanalysis relative to TOAR observations across North America, Europe, and East Asia. The global distributions of ozone bias predicted by ML revealed systematic patterns influenced by meteorological conditions, geographic features, anthropogenic activities, and biogenic emissions. The primary drivers identified include temperature, surface pressure, carbon monoxide (CO), formaldehyde (CH2O), and nitrogen oxide (NOx ) reservoirs such as nitric acid (HNO3) and peroxyacetyl nitrate (PAN). The ML framework provided a detailed quantification of the magnitude and variability of these drivers, delivering bias-corrected ozone estimates suitable for human health and environmental impact assessments. The findings provide valuable insights that can inform advancements in chemical transport modeling, DA, and observational system design, thereby improving surface ozone reanalysis. However, the complex interplay among numerous parameters highlights the need for rigorous validation of identified drivers against established scientific knowledge to attain a comprehensive understanding at the process level. Further advancements in ML interpretability are essential to achieve reliable, actionable outcomes and to lead to an improved reanalysis framework for more effectively mitigating air pollution and its impacts.

本研究采用可解释机器学习（Explainable Machine Learning, ML）框架，探究全球化学再分析中近地面臭氧偏差的区域依赖性及其潜在驱动因素。本研究采用2005—2020年对流层臭氧评估报告（Tropospheric Ozone Assessment Report, TOAR）观测网络的近地面臭氧观测数据，以及多模式多组分化学（Multi-Model Multi-Constituent Chemical, MOMO-Chem）数据同化（Data Assimilation, DA）系统产出的同期化学再分析结果，开展机器学习训练。基于回归树的随机集成机器学习方法，成功复刻了北美、欧洲与东亚区域内，化学再分析相对于TOAR观测的臭氧偏差时空分布特征。机器学习预测得到的全球臭氧偏差分布，呈现出受气象条件、地理特征、人为活动与生物源排放共同影响的系统性特征。本研究识别出的核心驱动因素包括温度、地面气压、一氧化碳（Carbon Monoxide, CO）、甲醛（Formaldehyde, CH₂O），以及氮氧化物（Nitrogen Oxide, NOₓ）储库，如硝酸（Nitric Acid, HNO₃）与过氧乙酰硝酸酯（Peroxyacetyl Nitrate, PAN）。该机器学习框架可对上述驱动因素的强度与变异性进行精细化量化，生成经偏差校正的臭氧估算结果，适用于人体健康与环境影响评估相关工作。本研究结果为化学传输模型、数据同化以及观测系统设计的优化提供了宝贵参考，进而助力近地面臭氧再分析精度的提升。然而，众多参数间存在复杂的相互作用，这表明需结合已有科学认知对识别出的驱动因素开展严格验证，才能在过程层面实现全面认知。进一步优化机器学习可解释性，对于获得可靠且可落地的研究结果、完善再分析框架以更有效地减缓空气污染及其影响至关重要。