Intelligent Corrosion Evaluation and Prediction via Data-driven Approaches

Dawei Zhang^1,2*, Tao Yang^2,3, Dongmei Fu^2,3, Xiaogang Li^1,2

¹ Beijing Advanced Innovation Center for Materials Genome Engineering, University of Science and Technology Beijing, Beijing 100083, China

² National Materials Corrosion and Protection Data Center, Beijing 100083, China

³ School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China

EXTENDED ABSTRACT: Materials corrosion is a complex failure process affected by many factors and can seriously jeopardize the service safety of infrastructure and major equipment. The estimated cost of corrosion losses adds up to more than 3 trillion yuan in China every year. The evaluation of corrosion resistance and behaviors takes the longest time and the highest cost in the whole chain of materials R&D and application. As such, corrosion evaluation is considered as one of the most challenging topic to study in the U.S. Materials Genome Initiative and is also a key and featured research direction in the field of material genome engineering in China.

Materials corrosion processes are often described as high-dimensional functions as they depend on many parameters including materials physical and chemical properties and environmental factors and also their complex interactions. To efficiently understand the relationship between high-dimensional corrosion data and influencing parameters, it is necessary to reduce the dimensionality of corrosion data and select some key characteristic parameters; Then the mapping relationship between corrosion data and key characteristic parameters can be established. Based on the multi-source and heterogeneous data of the National Materials Corrosion and Protection Data Center, this study adopts efficient machine learning methods, with comprehensive consideration of the source, structure and internal logical relationship of the corrosion data  and the application of different linear, local linear and nonlinear analysis methods. In views of the statistics, trend and combination characteristics, the influence parameters of materials corrosion are effectively selected and the data dimensionality is reduced via correlation and causality analyses. Subsequently, the nonlinear model between the key characteristic parameters and the corrosion process is constructed by using support vector machine, random forest and their optimized algorithms, in order to realize data-driven intelligent corrosion evaluation and prediction.