Structure And Properties Prediction Of New Superhard Materials Based On Machine Learning

Superhard materials are an important material with special functions in industrial production as well as in the electronics industry,etc.Currently,commonly used superhard materials such as diamond and artificial boron nitride have the disadvantage of high cost,so other alternatives need to be found.Although the traditional experimental trial-and-error method can also be effective in discovering new materials with higher hardness,it is difficult to keep up with the rapid development of the materials field today because of its high cost and time consuming disadvantages.With the gradual development of computer science,machine learning algorithms have been used for the design and property study of new materials to find new superhard materials that can efficiently obtain new materials that meet the conditions,thus reducing the time and cost required for the design of new materials,but machine learning algorithms for the property study of new materials to achieve high prediction accuracy has been a research challenge.In this paper,based on the machine learning algorithm and the collected material data,a series of regression models are developed to predict the elastic modulus and energy band information of the material,which not only can reduce the time and cost,but also improve the prediction accuracy,which can help to design new superhard materials for the corresponding applications and provide strategies for designing other functional materials.The hardness values of the material are calculated from the volume and shear moduli by equations,but it takes a long time to obtain these two modulus values using traditional methods,so three machine learning models are chosen to predict these two moduli,namely support vector machine regression,random forest and artificial neural network models,in order to shorten the research period.The data for training the models were obtained from the materials database website Matminer,and the cell atomic number(nsites),space＿group,volume,density and 132 magpie features obtained by expanding the molecular composition of the materials were selected as the feature descriptors in the study.The random forest model was found to have optimal results,but there were still some gaps with the expected results when making property predictions,so the model was re-optimized by adding new data,and its evaluation scores on the test set were RMSE of 6.626(volume modulus)and 6.774(shear modulus),and R~2 of 0.94(volume modulus)and 0.933(shear modulus),respectively.Using the optimized model to predict the bulk and shear moduli for a variety of materials,the predicted results were very close to the DFT theoretical values and their hardness were predicted by empirical formulas,with differences of about 10%from the hardness values obtained by other theoretical approaches,and the predicted hardness values were closer to the experimental values than those of others’models.In the study of the physical properties of superhard materials,the energy band information(band gap value)is a major focus of the study.The size of the band gap is significantly underestimated when using the traditional DFT method,and using the hybridized generalized HSE06 will improve the accuracy of the band gap,but its calculation time will also increase,even up to 106 seconds.Based on this,a neural network model(ANN)was chosen to train the model data to predict the elastic modulus,with new data obtained from the Materials Project database and the researched literature,and the final ANN performance score on the test set was 0.665for RMSE and 0.804 for R~2.Then compared with the theoretical value of HSE06,and found that the difference between them is concentrated around 15%,indicating that the model has a good prediction performance and can be used to predict the energy band information of other unknown materials.The combination of the above work shows that machine learning and first-principles calculations substantially improve the efficiency of material computational design and property prediction,and provide a theoretical basis for promoting the integration of machine learning techniques with the materials industry to meet engineering needs.