Material Design And Industrial Optimization Based On Machine Learning

Since‘Materials Genome Initiative’was proposed by the government of America in 2011,more and more attentions have been focused on the materials design and optimization based on machine learning.Screen and design of novel materials can be realized by using machine learning methods,in purpose of accelerating the process of novel materials development.Machine learning methods can also be used in the industrial optimization,where the key technical parameters may be controlled to optimize the quality and quantity of products.In this work,data mining（machine learning）methods were utilized to investigate the materials design of layered double hydroxides（LDHs）and ABO₃ perovskite manganites as well as the industrial optimization of fluorinated ethylene propylene（FEP）.The main achievements can be summarized as the following:（1）Materials design of LDHs based on machine learningIt is necessary to design the LDHs materials with big layer spacing in the application of electrode material for supercapacitors.In this work,the LDHs materials with bigger layer spacing(d_spacing)were screened out by using machine learning model.The data set consisting of 85 LDHs with d_spacing was collected from the published references.The machine learning model call XGBoost was constructed by using atomic parameters as descriptors.It was found that the related coefficient（R）and the mean relative error（MRE）of leave one out cross validation（LOOCV）for XGBoost model were 0.94 and 3.26%,respectively.The XGBoost model was shared on the online computation platform for materials data mining（OCPMDM）for the researchers.Through high-throughput screening,the novel LDH,Co_0.67Fe_0.33[Fe（CN）₆]_0.11·（OH）₂,was predicted with the biggest layer spacing of 12.40（?）,which is 10.91%larger than the biggest one（11.18（?））available in the training set.（2）Materials design of ABO3 perovskite manganite based on machine learningABO₃Perovskite manganites is a kind of excellent memory material because of the giant magnetoresistance effect and anisotropic magnetoresistance of antiferromagnetic materials.In this work,the ABO₃perovskite manganites with Néel temperature（T_N）larger than room temperature was screened out by using machine learning.The data set consisting of 159 perovskite manganites with T_N was collected from the published references.The machine learning model call support vector machine（SVR）was constructed by using the weighted averages of atomic parameters of doped A site as descriptors.It was found that the R and the MRE of LOOCV for SVR model were 0.87 and 18.67%,respectively.The SVR model was shared on the OCPMDM for the researchers.Through high-throughput screening,the novel ABO₃ perovskite manganite,Sr_0.7Ce_0.1Sm_0.2Mn O₃,was predicted with the highest T_N of 305.5K,which is5.34%higher than the highest one（290K）available in the training data set.（3）Industrial optimization of FEP based on machine learningMelt index is an important indicator of FEP,which can be affected by many technical parameters of polymerization process.Machine learning can be used to construct the relationship between the melt index of FEP and the technical parameters of polymerization process.Thus,the fluctuation range can be controlled based on the optimization of technical parameters.In this work,the melt index of FEP and the technical parameters of polymerization process were collected in the factory of FEP products.Firstly,feature selection was carried out by analyzing the importance of descriptors.The key factors affecting on index of FEP were found.Then,the principal component analysis（PCA）method was used to construct the optimal zone describing the ranges of optimal FEP products.At last,through two rounds from model building to model application,the ratio of samples with the melt index of FEP between 24±2fluctuation range was lifted from 22%to 53%.