Research On Screening Of Perovskite Functional Materials Based On Deep Learning And Transfer Learning

Materials are the cornerstone of technological development.However,the development and application of new materials all need to go through a long process.On the one hand,the traditional experimental-based material R&D model has difficulties such as high cost,long cycle,and performance that cannot meet application requirements.On the other hand,there is a huge unexplored material design space that can achieve more superior material performance.Since the introduction of the Material Genome Project,it has received widespread attention from materials research scholars.Its purpose is to achieve a 50%reduction in research and development costs and a time reduction of new materials through key technologies such as high-throughput calculation methods and material big data.Therefore,in order to accelerate the exploration of new materials,three main research directions have emerged:first,high-throughput material experiments;second,high-throughput material calculation based on first-principles and molecular dynamics;third,machine learning and Material genomics with deep learning for material performance prediction and reverse material design.In recent years,materials researchers have used machine learning methods to find laws in the historical data of materials and give guidance suggestions to improve the traditional"trial and error method",which has become a hot research direction in the field of materials.At present,as one of the most studied materials in the field of materials,perovskite has many excellent properties,so it has a broad application prospect.It has great potential in the discovery of new perovskite functional materials such as solar panels,superconductors,thermoelectric and catalytic materials.In this paper,one of the key obstacles to discovering new materials based on machine learning,that is,the lack of sufficient training data,proposes a transfer learning method based on perovskite structure information,using a small amount of perovskite structure information and element descriptors Train a high-precision machine learning model;then use the expanded data set to train a convolutional neural network model with strong predictive ability using statistical features of Magpie elements for screening ABX₃materials;and finally use the tolerance factor?to verify the selected ABX₃ Whether the material is perovskite functional material.The specific research work is as follows:First,in the study of material attribute prediction,all data samples obtained from the database are usually used to train the prediction model.However,the high redundancy of the material database samples leads to the strong bias and overfitting of the trained model.In view of this situation,this paper proposes a feature selection method for correlation and redundancy analysis to select representative sample data from the data set.Secondly,this paper calculates the structural characteristics and element characteristics of a few perovskite with structural information through Pymatgen,a total of 31 feature descriptors.Using these 31 features training to obtain a high-precision transfer learning model,this model uses the gradient lifting regression algorithm in the machine learning algorithm.Then use this high-precision transfer learning model to predict perovskite materials without structural information.The experimental results show that the 31 mixed feature descriptors calculated in this paper predict that the formation of perovskite is superior to Ong?Descriptors and Magpie descriptors proposed by other researchers.And in the commonly used machine learning models,the gradient lifting regression used in this paper is better than the prediction effects of the models such as random forest regression,Lasso,and support vector machine regression.The transfer learning model combined with the hybrid feature descriptor proposed in this paper can solve the problem of small data sets for discovering new materials based on machine learning.Secondly,in order to screen out stable ABX₃ materials,this paper proposes a screening model based on convolutional neural networks.Label ABX₃ materials without structure through a high-precision transfer learning model;then calculate their Magpie element statistical feature descriptors.Magpie calculations do not require structural information;Finally,the convolutional neural network model and the data represented by Magpie descriptor Training together,you can get a general ABX₃ material screening model.By comparing this model with Elem Net model and machine learning model,the prediction effect of the convolutional neural network model proposed in this paper is the best.Through the convolutional neural network model,we selected ABX₃ materials with a formation energy of less than 0 from 21316 ABX₃materials.Finally,this paper verifies whether the selected ABX₃ material is a stable perovskite material through a new tolerance factor?,and?can accurately determine whether the selected ABX₃ material is perovskite or non-perovskite.Verification of perovskite materials with?requires only chemical composition,which makes many perovskite materials with unknown structure hope to be verified.In addition to predicting whether a material is a stable perovskite material,?also provides a monotonic estimate of the probability that a material will be stable in a perovskite structure.After a large number of calculation experiments,this paper has selected promising new perovskite materials from 21,316 hypothetical ABX₃ materials.Some of these stable perovskite materials have been reported in other literatures.The rest are subject to further experiments or DFT calculations.