Machine Learning Assisted Design And Experimental Verification Of Ultra-fine WC-Co-based Cemented Carbide

Cemented carbide is widely used in high-speed cutting tools,exploration drilling,mining and other fields because of its high hardness,high strength,good wear resistance and corrosion resistance.By adding an appropriate amount of refractory transition metal carbides or nitrides into the hard phase,the grain growth can be restrained,and the ultra-fine cemented carbide with high comprehensive mechanical properties with high hardness and strength can be prepared.The work of this paper aims at the key problems in the field of ultra-fine WC-Co cemented carbide at present,utiltizing the data-driven research ideas and machine learning algorithms to explore the relationship between composition,structure and performance,as well as proposing corresponding solutions,providing guidance for the forward and reverse design of alloys.The main research contents and innovations of this paper are as follows:(a)As an interdisciplinary field,the application of machine learning technology is time-consuming for material researchers who lack the professional background of data science and computer programming skills.In view of this problem,Data Metallurgy software wich integrating feature engineering,machine learning,hyperparametric optimization,performance prediction and other functions was developed.The software provides a simple and easy-to-use graphical operation interface,which provides an alternative scheme for material researchers to rapidly apply machine learning technology into practical research.(b)For the "big data of materials" problem,the amount of research data of ultrafine cemented carbide is still limited,and there are many key properties indicators to be considered in designing high-performance ultra-fine cemented carbide.In response to this problem,this work proposes a multiple loops machine learning framework to predict the properties of WC-Co based cemented carbides.Through several cycles of"data acquisition-feature engineering-machine learning-model prediction",scientific cemented carbide alloy performance database and key performance machine learning model library are obtained at the same time,providing the basis for data and model for efficient alloy design in the future.(c)As a brittle material,the common reliability evaluation method of cemented carbide is to combine Weibull statistics,but there are several problems such as unequal or insufficient data,incomplete identification trend,and errors between different methods,which limit the application of reliability analysis in actual production.For addressing this issue,this work utilized the limited alloy strength data series to design and verify the proposed universal data mining aided alloy reliability analysis method,which provides scientific guidance for the process parameters and reliability design in actual production.(d)At present,the alloy design of empirical method is costly and inefficient,and the alloy design of computational materials science method is slightly insufficient in the optimization of property indicators and scope of applications,while most machine learning algorithms,as "black box" models,are difficult to reversely predict the composition based on performance indicators,and the traditional high-throughput screening method consumes too much computational resources and is inefficient.In order to solve this problem,this work designed the Bayesian methods assisted principal component reverse prediction strategy.By limited the search range through t-SNE and principal component analysis model,the alloy components meeting the required performance indicators were obtained with a combination of Bayesian method and reverse solving the principal component analysis model,and verified in the ultra-fine WC-Co cemented carbide system.As an efficient material design tool,this method provides a new perspective for property-oriented alloy design and shows a practical application prospect.