Study On Prediction Of Nucleation Undercooling Degree And Grain Size Of Alloy Via Machine Learning

Research and development of materials play a fundamental role in advancing technology and upgrading industry.The traditional‘stir-fry'trial-and-error method,used to develop advanced materials,usually leads to high cost of time,human resource and budget.What is worse,it can not meet the demand about advanced materials of industrial application since service environment becomes more complex and various.With the rapid development of the database and the data mining,Machine Learning has shown huge potential in accelerating materials design and development.By employing large amount of data generated from scientific experiments,simulations,etc.,researchers have induced machine learning to speed up materials research,leading to great success in materials.This thesis uses machine learning to carry out data-driven prediction of undercooling degree and grain size for irons and aluminium alloys during casting.All the data are collected from previous studies,later analyzed by machine learning method including correlation analysis,feature selection,regression analysis,and integrated algorithm.The calculation formula of Bramfitt mismatch can be modified and optimized to accurately characterize the interface mismatch of any crystal structure.Based on machine learning,this thesis predicts the undercooling degree of irons and aluminum alloys during casting,and established a data-driven quantitative prediction model.And then the grain size of Al and Al-7Si after casting is predicted by four integrated algorithms.The main results are as follows:(1)This paper is based on Bramfitt mismatch equation for correction and optimization,The relationship between undercooling degree and interface mismatch degree under different equation is compared.The results show that the two new methods can accurately characterize the interfacial mismatch of any crystal structure without judging the substrate and nucleation phase,and the mismatch value is unique.The R~2of Bramfitt method and the proposed method are 0.61,0.68 and 0.76,respectively,indicating that the proposed method has a better fitting degree and is more applicable.(2)Nine machining learning models,including Random Forest,e Xtreme Gradient Boosting(Xgboost),Ridge Regression(RIDGE),and Gradient Boosting Regressor method are used to predict the undercooling degree via six features,which include the cooling rate,mean atomic covalence radius,and mismatch.Then,four additional effective models of machine learning algorithm are selected for further analysis and cross-validation.Finally,the optimal machine learning model is selected for the data set,and the best combination of features is found by comparing the prediction accuracy of all possible feature combinations.It is found that the Random Forest model with the cooling rate(CR)and mean covalent atomic radius(MAR)features has the optimal performance results for predicting the undercooling degree.(3)Four machine learning models including Extreme Gradient Boosting(Xgboost),Random Forest,Ada Boost,and Gradient Boosting Decision Tree(GBDT),are used to establish relationships between the grain size of Al and Al-7Si and the factors such as alloy composition,grain refiners and processing parameters.The results show that the Xgboost algorithm is the best model with a coefficient of determination R~2of 0.90,mean absolute error of 43.84,and root mean square error of6.56.It concludes that the composition of the grain refiner and Ti are the key factors which could predict the grain size effectively.