The Research Of Materials Genome For Perovskite Functional Materials

Material Genome Initiative(MGI)has been attracting a lot of attention from material researchers since 2011.The MGI develops an infrastructure to shorten the materials development cycle by using high-throughput computation,high-throughput experimentation and data mining techniques for materials.In recent years,data mining techniques have been widely used in materials researches.It is a hot topic to find the "discipline" of the materials and provide experiment guidance from history materials data by using machine learning algorithms in materials researches.The MGI methodology can be utilized to reduce experimental cost and development time instead of the traditional "trial and error" method.However,it is very difficult for materials researchers to construct and apply machine learning models without a convenient and practical platform for materials data mining.In order to avoid additional learning costs from materials researchers,we developed an online machine learning platform(called AMINER)for those who are unfamiliar with data science.AMINER provides classification,regression,matrix decomposition,feature selection and hyperparameters optimization algorithms,which are widely used in materials machine learning tasks.Also,automatic generation of molecular descriptor and model sharing are tailor-made for perovskite materials.By using AMINER,researchers can rapidly and conveniently build and optimize a variety of different machine learning model for regression or classification.Based on the machine learning model,the virtual screening can be carried out to screen out a candidate material with designed property within a given range,which is helpful to further experiments.In addition,deep feature network(DFN)was proposed for the feature selection of datasets of perovskite ferromagnetic Curie temperature based one deep neural networks.The performance of DFN was compared with those of the principal component analysis(PCA),partial least squares(PLS)and genetic algorithm(GA).By cross-validation,DFN showed the advantages in feature processing compared with three common feature processing algorithms both in accuracy and robust.Furthermore,for obtaining better performance in dataset with small size,the DFN trained by ferromagnetic Curie temperature datasets was used as the pre-training model in ferroelectric Curie temperature.Using fine-tuning,the DFN exhibits good performance in modeling of ferroelectric Curie temperature.To provide better service of materials data mining,we designed an intelligent modeling process that was implemented in the platform of AMINER.The model constructed via the intelligent modeling process,shows comparable performance in both classification and regression tasks in ferromagnetic Curie temperature dataset,compared with GA-SVM methods.At the same time,the intelligent model process is efficient in model building,it only takes half time of using GA-SVM.The intelligent modeling process is suitable to build preliminary model for material machine learning tasks.High-throughput screening is one of the main applications using material machine learning model.In this work,virtual screening based on genetic algorithm or efficient global optimization was designed.Compared with the traditional Monte Carlo algorithm,the GA virtual screening shows efficient screening speed.The screening time can be shortened to 1-2 minutes by GA virtual screening instead of Monte Carlo algorithm spending several weeks.For convenience of materials researchers,GA virtual screening has been integrated in AMINER.Material researchers only need to set the screening material range and specify the model to start virtual screening.EGO using prediction means and standard deviation improves model prediction with a more reliable screening result,because the traditional method screens the results without the prediction confidence.