A hybrid knowledge-guided and data-driven method for predicting low-alloy steels performance

Abstract

Low-alloy steels are essential engineering materials and precise prediction of their performance is crucial. Data acquisition in the steel industry is costly, and the available data volume is usually small. So far, with limited data, both traditional material modeling methods and data-driven machine learning cannot accurately build the relationship among material composition, process and performance. Although there is a wealth of experience and knowledge accumulated through extensive research in the steel field, combining the knowledge with data remains a significant challenge. To fully utilize domain knowledge, in this paper, a knowledge-driven graph convolutional network (KD-GCN) method is proposed for predicting the performance of low-alloy steels. Firstly, a knowledge graph is constructed by using steel knowledge from professional books. Then structured steel composition and processing data are transformed into a graph representation. Next, through a multi-layer graph convolutional network, the domain knowledge and structured data are fused to predict the strength and plasticity of steels. Furthermore, the proposed method is tested on a low-alloy steel dataset. Comparison results demonstrate that KD-GCN outperforms some other machine learning methods without using domain knowledge. Finally, feature importance analysis experiments show that the obtained influence of composition and process on steel performance is highly consistent with domain knowledge, which further validates the effectiveness of the proposed method.