Property Prediction of Maraging Steels: Machine Learning vs. Physical Metallurgical Modelling

Chi Zhang¹*, Chunguang Shen², Chenchong Wang², Wei Xu²

¹ Key Laboratory of Advanced Materials of Ministry of Education, School of Materials Science and Engineering, Tsinghua University, Beijing 100084, China

² State Key Laboratory of Rolling and Automation, Northeastern University, Shenyang, Liaoning 110819

ABSTRACT:  The hardness prediction models of maraging steel were selected to compare the characterises between Physical metallurgical (PM) and Machine Learning (ML) models. Two PM models referred as WMOS and GRR for yield strength were adopted while SVR and RF were used in ML models. Dataset D1 was collected from the literatures on maraging steels strengthened by eight sorts of precipitates and D2 was based on another precipitate kind R-phase. D3 was the union set of D1 and D2 to study the generation ability. Three modelling strategies for PM models were applied: (1) all the parameters were extracted from literatures; (2) only some uncertain parameters were fitted; (3) all important tuning parameters were fitted. The upper limit of model prediction error drops in order of models with strategy (1), (2) and (3). However, the best results of PM models are greatly worse than that of ML models. PM model is sensitive to the alloy system, while ML model is more sensitive to data quality. The prediction based on the ML model is still a ‘black box problem’ due to the lack of microstructural information, representing that it is less analysable than the PM.

Figure 1. The comparative results of PM and ML methods in terms of predicted accuracy