The Developments And Applications Of The High-throughput Computational Methods And Toolkits Combining With Machine Learning For Materials Design

For the development of first-principles density functional theory,the materials design methods including high-throughput computational methods,crystal structure prediction methods are playing increasingly important roles.First principles high-throughput calculations allow large-scale searches of material searching space to discover new materials,properties,and principles.In recent years,with the breakthrough of artificial intelligence image recognition and other technologies,machine learning algorithms have been rapidly applied in many different disciplines.The intercross and integration of machine learning and other material design methods has made remarkable progress in the discovery of new materials,new structure-property relationships,design principles,and so on,which further stimulates the continuous attention of many scientific researches.The effective integration of material design methods such as the combination of high-throughput computational methods with machine learning algorithms depends on the development of new algorithms and software infrastructure.How to efficiently generate,collect,manage,learn and mine large-scale material data are all the main difficulties in developing such algorithms and software.To solve these problems,we have developed three computational methods and software based on high-throughput computational methods and machine learning algorithms,and applied them to the study of the physical properties of some typical optoelectronic semiconductors.The following innovative achievements or progress have been made:(1)Developing the crystal structure data reading and writing modules and structure prototype database application programming interface for the software called“Jilin Artificial-intelligence aided Materials design Integrated Package”(JAMIP),and running 1000-level high-throughput computing task tests.There are many different file formats of crystal structural materials.Correctly reading/writing these files is one of the import task in the development of materials design software.We have designed and developed reading and writing algorithms for material structure files with different file format.Especially for the various nonstandard formats of Crystallographic Information Files(CIF),which can't be read and analyzed by other released algorithms and software,we have designed and developed new adaptive algorithms to process these different types of nonstandard CIF files to ensure wider universality of the structure reading and writing modules.We have developed a structure prototype database application programming interface and supporting tools for JAMIP software to facilitate high-throughput structure modeling and calculation.In addition,we have carried out 1000-level high-throughput computing task tests on JAMIP to verify the reliability of this software.(2)Developing a set of structure prototype generation software called“Structure Prototype Generator Infrastructure”(SPGI)based on artificial intelligence classification algorithm,and using it to construct a large inorganic crystal structure prototype database called“Local Atomic Environment based Inorganic Crystal Structure Prototype Database”(LAE-ICSPD).Crystal structure prototypes are essential while the virtual materials as the initial entries for date-driven or high-throughput computational materials design frameworks are usually generated by decorating the prototype structure frames using varying combinations of atoms.Therefore,the high data quality and high uniqueness of the structure prototypes are essential for the high-throughput computational materials design frameworks.Besides,the local atomic environments(LAEs)can effectively and entirely describe the atomic configuration and uniqueness of crystal structures.We have developed an artificial intelligence crystal structure prototype generation software SPGI,which is based on unsupervised learning strategy and uses the local atomic environment of crystal structure as descriptors to do clustering analysis for all the inorganic crystal structures synthesized experimentally.Finally,15613 structural prototypes have been selected from it,based on which,a large inorganic crystal structure prototype database called LAE-ICSPD has been constructed.It can provide the required crystal structure prototype data for material design methods such as high-throughput computational materials design methods or machine learning methods.(3)Developing a new structure characterization method from which,the crystal structures can be inversely constructed,and designing a strategy of inverse crystal structure prediction based on high-throughput computational materials design methods and machine learning algorithms.Based on the crystal structure projection decomposition algorithm,we have developed a new structure characterization method called“two-dimensional slice lattice graph”,which can be used to inversely reconstruct the crystal structure.The core idea is that slice and project the three-dimensional crystal structures.The atoms belonging to the same plane will be projected into the same two-dimensional mesh grid graphs.This descriptor can be used as the input for both a supervised learning prediction model(such as deep neural network)and an unsupervised learning generation model(such as autoencoder and generative adversarial network).A new inverse crystal structure prediction strategy have been designed,which is guided by the target properties of materials,combining with the crystal structure prototype database,supervised learning property prediction model and unsupervised learning generation model.(4)Combing the experimental and theoretical calculations,two different surface structures of Cs Pb Br₃ perovskite experimentally observed by scanning tunneling microscopy(STM)have been identified.The mechanism of the mutual transformation between the two surfaces has also been explained.Two different stable surfaces of inorganic perovskite Cs Pb Br₃ were experimentally observed by STM,named the"stripe"surface and the"armchair"surface.The"stripe"surface area is larger than the"armchair"surface area.However,experimenters cannot know the specific surface atomic arrangement of these two surfaces and why the“stripe”surface area is larger.We used JAMIP to run the high-throughput calculations of STM simulations for Cs Pb Br₃ perovskite system.Finally,we successfully found two slab phases,whose simulated STM images of surfaces can match the experimentally observed images and then resolved the specific surface atomic arrangement of these two surfaces:"stripe"originates from the long and short distance spacing of surface Br pairs and Cs atoms beside the pair;"armchair"originates from the hierarchical arrangement of Br pairs and beside Cs atoms in the surface.By calculating the surface energy of these two slab phases,we found that the surface energy of the"armchair"surface is slightly higher than that of the"stripe"surface,that is,the"stripe"surface is more stable,therefore the"armchair"surface tends to spontaneously transform to the"stripe"surface,herein,a larger area of the"stripe"surface area was experimentally observed.(5)Performing the first principles based high-throughput calculations to investigate why doping 5-AVA into pure MAPb I₃ perovskite can effectively improve the material's stability,and how the band gap values and the electron mobilities vary with the change of the number of atomic layers for?-phase and?-phase of two-dimensional layered indium selenide materials.For organic perovskite MAPb I₃ materials,experimenters found that doping 5-AVA into its pure phase to get(5-AVA)_xMA_1-xPb I₃ phase could significantly improve its stability under various complex conditions.We performed high-throughput calculations to get tens of MAPb I₃and(5-AVA)_xMA_1-xPb I₃ phases'formation energies.Finally,two MAPb I₃ phases and two(5-AVA)_xMA_1-xPb I₃ phases with the lowest formation energies were retained for further calculations.After analyzing the decomposition enthalpies and the degree of octahedron distortion of these four phases,we found that the bonding between 5-AVA⁺ions and I^-ions was stronger(resulting in greater octahedron distortion).The decomposition enthalpies of(5-AVA)_xMA_1-xPb I₃ systems were also lower.It was inferred that the addition of 5-AVA made it more difficult for the organic molecules in the organic perovskite to escape.This is the main reason why doping 5-AVA into pure MAPb I₃ perovskite can effectively improve its stability.Moreover,by developing high-throughput computational processes and modules for electronic transport calculations with collaborators,we studied how the band gaps and electron mobilities vary with the number of atomic layers in two-dimensional layered indium selenide materials(?-phase and?-phase).We found that the band gap values of both phases decrease with the increase of the number of atomic layers,and the electron mobility of both phases increases with the increase of the number of atomic layers.