The Exploration Of Growth Of Single Crystal By Flux Method By Machine Learning Method

The growth of high-quality single crystals is of great significance for the study of condensed matter physics.Single crystals are vital prerequisite for extensive scientific research fields,such as condensed matter physics,surface science,lasers and nonlinear optics.Fundamental studies like QHE/FQHE,Wyle semi-metal,etc.,all relied on the production of high-quality single crystals.The exploration of suitable growing conditions for single crystals is expensive and time-consuming,especially for ternary compounds because of the lack of ternary phase diagram.The method of crystal growth by flux is a very important method of crystal growth,and it is of great significance to the study of crystal properties.In theory,as long as a suitable flux or flux combination for a single crystal can be found,the single crystal can be grown by flux method,and the successful growth conditions are repeatable.Using flux method to grow single crystals,the choice of flux and the determination of growth conditions are very important.In this paper,the experimental data of single crystal growth with fluxes in laboratory are used to find the conditions of crystal growth by means of machine learning.Specific research is as followsHistorical growth data are collected from cooperation groups,which is defined as Group I and Group II,respectively.and our Group I basically contain all the elements in the periodic table.Our research is mainly focused on Group I,Group II as a discussion of the consistency of machine learning in this issue.Based on the experience of single crystal growth by flux method,the growth conditions,the physical and chemical properties of fluxes and solutes which may affect the crystal growth are selected.Growth conditions include maximum temperature,centrifugal temperature,temperature difference,cooling rate,high temperature residence time,flux type,raw material type,raw material content.The physicochemical properties of flux and solute include melting point,vapor pressure,atomic number,relative atomic mass,electronegativity negativity,density,phase diagram.These attributes are regarded as the features of machine learning training,and the growth results are labeled.Four machine learning algorithms,decision tree,random forest,support vector machine and gradient boosting decision tree,are used to find the optimal model.SVM method outstands others,with 81% accuracy predicting experimental results on the test set of group I.Laboratory accuracy,in contrast,is only 36%.average accuracy rate of the model is 83%,the average F1 fraction is 83%,the precision rate of successful samples is 78%,and the recall rate is 66%.Compared with the experimental data,the artificial precision rate is 30%,and the successful growth rate of the model selected by machine learning training is 49% higher than that by artificial selection.The accuracy and F1 score of random forest are the highest in the three models,and the precision of SVM model is the highest in the three models.Changing the threshold of models can change the precision of successful samples,so we can adjust the threshold to satisfy our application requirements.With continuous growing of the database,the ML will be better at the prediction of single crystal growth and uncover more essential rules in the process.The machine learning model of single crystal growth by flux method can be used in the laboratory to help the laboratory grow single crystal better.This paper introduces the method of applying this model to the laboratory.The first method is that the model grows independently and traverses all possible feature combinations of a crystal.Then the trained model is used to fit these combinations.The probability of successful growth of these conditions can be obtained by fitting.A threshold can be selected and the conditions above the threshold probability are given priority according to the probability level.Long.The second method is that the model assists the growth of human experience.The experimenter selects the conditions according to his own experience,and then judges the growth of the conditions.The decision tree model has good resolution.We find that the electronegativity of flux is an important physical property describing its function,and it is important to choose other conditions for crystal growth.The electronegativity of fluxes with less electronegativity requires less physical and chemical properties of the constituent elements of single crystals,while the fluxes with higher electronegativity require more physical and chemical properties of the constituent elements of growing crystals.If electronegativity is needed for growth,different crystals of different physical and chemical properties require different fluxes.The effect of cooling rate and variance density on the final crystal growth is significant.Moreover,we found that the lower the cooling rate,the closer the flux density and the element density of the crystal,the easier the crystal to grow successfully.Therefore,the tree graph can show some factors that are often overlooked or less considered in laboratories.We have found several rules are summarized from the decision tree,which are divided into two types.One that already has are well understood a clear theory,but because there are many factors should be combined with,is easily ignored but re-discovered by ML.And the other has no clear theory.The former includes such as:(a)When the cooling rate is low,there are fewer factors need to be considered.Otherwise,there are more factors have to be considered.better single crystals growth is associated with(a)lower cooling rate and(b)Lower melting point of B in flux is more likely to succeed.ML also discovers some rules without clear explanatory theory,which are of particular interest.For example,ML suggests that it is difficult to grow single crystals if:(a)the x/z and y/z values are relatively large;(b)the average difference between the liquid phase density of flux and the components of single crystal(A,B,and C defined above)is large.Furthermore,it is recommended to choose a flux with a smaller maximum electronegativity differences between A,B and C.