4)National Materials Corrosion and Protection Data Center,Beijing 100083,China 5) School of Automation and Electrical Engineering,University of Science and Technology Beijing.Beijing 100083,China Corresponding author,E-mail:fdm_ustb@ustb.edu.cn;dzhang@ustb.edu.cn ABSTRACT Machine learning algorithms are widely used to predict the corrosion rate of materials in a specific environment,but such black-box models'interpretability are poor,which obstacle the application in the field of material corrosion.Therefore,to increase the algorithm transparency in practical applications,it is necessary to further explore the causal relationship in the material corrosion phenomenon based on machine learning models.To solve the above problems, this paper aims at the corrosion process of carbon steel in the atmosphere with many variables and complex mechanisms, proposing an important variable mining framework based on the comprehensive intelligent model.This framework can mine the important environmental variables that affect the early atmospheric corrosion of carbon steel and their influence on the corrosion galvanic current.This paper collected hour-level atmospheric corrosion data of carbon steet 45 from 5 test sites in China through the atmospheric corrosion monitor (ACM)and China Meteorological Administration,including relative humidity,temperature,rainfall,O;concentration,SO2 concentration,NO2 concentration.PM2.5.concentration.PM10 concentration.In order to ensure the stability of the results,three machine learning models with different fitting strategies are constructed:Random Forest (RF).Gradient Boosted Regression Trees (GBRT),and Back-Propagation Neural Network (BPNN).Then it is important to use Multi-model Ensemble Important Variable Selection (MEIVS)to quantify the importance of environmental variables and extract important environmental variables that affect the early atmospheric corrosion of carbon steel severely.Eventually,the Partial-dependence Plot (PDP)between environmental variables and corrosion galvanic current is drawn.Based on the simulation results,three significant conclusions are obtained.1)Compared with Pearson's Correlation Coefficient (PCC)and Spearman's Correlation Coefficient(SCC),the important environmental variables mined by MEIVS algorithm are more in line with the priof law of early atmospheric corrosion of carbon steel. Relative humidity,temperature,and rainfall have the greatest impaot on the early atmospheric corrosion of carbon steel,and O also has a great influence on the atmospheric corrosion in Sanya.In addition,other pollutants in various regions have a weak impact on the early atmospheric corrosion of carbon steel.2)PDP shows that in most cases,the corrosion galvanic current of carbon steel 45 is negatively correlated with temperature and positively correlated with relative humidity.3)PDP and MEIVS are well consistent.The simulation reveal that PDP corresponding to important environmental variables has a greater range of change,and the chariging t nd of PDP can reflect the influence of environmental variables on corrosion galvanic current. KEY WORDS atmospheric on steel;model integration;important variable extraction;partial-dependence plot 碳钢的大气腐蚀是一 种电解质膜下的电化学反应山。由于腐蚀现象的广泛存在，并且近年来对 腐蚀过程细致深的研究，促使人们对腐蚀模型不断提出了更高的精度要求，还希望通过模型可以 对腐蚀现象做出机理解释。相对湿度、温度以及降雨与液膜厚度的变化相关，在大气腐蚀过程中起 到重要作用2，常见的硫化物、氯化物、臭氧、固体颗粒物等污染性杂质对大气腐蚀也有促进作用 。影响腐蚀过程的环境变量众多，并且不同的环境变量之间也会产生相互作用，使得金属的大气 腐蚀更加复杂例。 传统的腐蚀挂片法通过把参数已知的金属试片放入特定环境下，计算暴露期间内试样的质量变 化来计算腐蚀速率。IS09226-2012标准给出了测定标准试样在大气环境下腐蚀速率的方法。IS0 9223-2012标准o则以相对湿度(Relative Humidity,RH)、温度(Temperature,.T)、SO2沉积速率、 C沉积速率的年平均值量化碳钢等四种材料在第一年的腐蚀速率，并据此推测环境的腐蚀性等级。 基于传统挂片法，支元杰叫利用失重法测得Q235碳钢在10个试验地点的腐蚀速率，结合随机森林4) National Materials Corrosion and Protection Data Center, Beijing 100083, China 5) School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China  Corresponding author, E-mail: fdm_ustb@ustb.edu.cn; dzhang@ustb.edu.cn ABSTRACT Machine learning algorithms are widely used to predict the corrosion rate of materials in a specific environment, but such black-box models’ interpretability are poor, which obstacle the application in the field of material corrosion. Therefore, to increase the algorithm transparency in practical applications, it is necessary to further explore the causal relationship in the material corrosion phenomenon based on machine learning models. To solve the above problems, this paper aims at the corrosion process of carbon steel in the atmosphere with many variables and complex mechanisms, proposing an important variable mining framework based on the comprehensive intelligent model. This framework can mine the important environmental variables that affect the early atmospheric corrosion of carbon steel and their influence on the corrosion galvanic current. This paper collected hour-level atmospheric corrosion data of carbon steel 45 from 5 test sites in China through the atmospheric corrosion monitor (ACM) and China Meteorological Administration, including relative humidity, temperature, rainfall, O3 concentration, SO2 concentration, NO2 concentration, PM2.5 concentration, PM10 concentration. In order to ensure the stability of the results, three machine learning models with different fitting strategies are constructed: Random Forest (RF), Gradient Boosted Regression Trees (GBRT), and Back-Propagation Neural Network (BPNN). Then it is important to use Multi-model Ensemble Important Variable Selection (MEIVS) to quantify the importance of environmental variables and extract important environmental variables that affect the early atmospheric corrosion of carbon steel severely. Eventually, the Partial-dependence Plot (PDP) between environmental variables and corrosion galvanic current is drawn. Based on the simulation results, three significant conclusions are obtained. 1) Compared with Pearson’s Correlation Coefficient (PCC) and Spearman's Correlation Coefficient (SCC), the important environmental variables mined by MEIVS algorithm are more in line with the prior law of early atmospheric corrosion of carbon steel. Relative humidity, temperature, and rainfall have the greatest impact on the early atmospheric corrosion of carbon steel, and O3 also has a great influence on the atmospheric corrosion in Sanya. In addition, other pollutants in various regions have a weak impact on the early atmospheric corrosion of carbon steel. 2) PDP shows that in most cases, the corrosion galvanic current of carbon steel 45 is negatively correlated with temperature and positively correlated with relative humidity. 3) PDP and MEIVS are well consistent. The simulation reveal that PDP corresponding to important environmental variables has a greater range of change, and the changing trend of PDP can reflect the influence of environmental variables on corrosion galvanic current. KEY WORDS atmospheric corrosion; carbon steel; model integration; important variable extraction; partial-dependence plot 碳钢的大气腐蚀是一种电解质膜下的电化学反应[1]。由于腐蚀现象的广泛存在，并且近年来对 腐蚀过程细致深入的研究，促使人们对腐蚀模型不断提出了更高的精度要求，还希望通过模型可以 对腐蚀现象做出机理解释。相对湿度、温度以及降雨与液膜厚度的变化相关，在大气腐蚀过程中起 到重要作用[2-4]，常见的硫化物、氯化物、臭氧、固体颗粒物等污染性杂质对大气腐蚀也有促进作用 [5-7]。影响腐蚀过程的环境变量众多，并且不同的环境变量之间也会产生相互作用，使得金属的大气 腐蚀更加复杂[8]。 传统的腐蚀挂片法通过把参数已知的金属试片放入特定环境下，计算暴露期间内试样的质量变 化来计算腐蚀速率。ISO 9226-2012 标准[9]给出了测定标准试样在大气环境下腐蚀速率的方法。ISO 9223-2012 标准[10]则以相对湿度（Relative Humidity, RH）、温度（Temperature, T）、SO2沉积速率、 Cl-沉积速率的年平均值量化碳钢等四种材料在第一年的腐蚀速率，并据此推测环境的腐蚀性等级。 基于传统挂片法，支元杰[11]利用失重法测得 Q235 碳钢在 10 个试验地点的腐蚀速率，结合随机森林 录用稿件，非最终出版稿