Theoretical Simulation Of Materials Combining First Principles Calculation And Machine Learning

For a long time,new materials have been the main factor to promote the development of material application and technology.Over the past decade,many advances in basic knowledge and technical applications in this field have involved the unexpected discovery of new materials with novel and desirable properties.Many of these materials have been transformed from being an object of basic research interest to being an object of practical application.However,the complexity of new materials has brought difficulties to the realization of experimental science,and many processes are difficult to capture,which limits the rational design of materials.With the enhancement of computer capabilities and the development of new calculation methods,theoretical calculation methods can be used to reveal the relationship between the structure and properties of these materials to explore new materials with excellent performance.Meanwhile,the combination of machine learning algorithms provides new ideas for the design of new materials.The properties of the materials can be effectively adjusted by adjusting the composition and surface morphology of the material.A deep understanding of the relationship between structure and effect can accelerate the design and development of new materials.The original work of this thesis is divided into two parts:we first studied various means of chemical modification on borophenes using first-principles density functional theory calculations.Particularly we examined tuning of band gap and stability of materials of interest,based on which we can predict new candidates for semiconductor materials.Theen,we have combined first-principles calculations and machine learning techniques to establish the quantitative relationship between carbonyl structure and dissociation energy,which is of great significance to the development of carbonyl electrode materials.This thesis consists of four chapters.The first chapter mainly elaborates the background and current situation of the two parts for the research direction and the problems that need to be solved.The first part introduces the current experimental science using molecular beam epitaxy(MBE)technology to successfully synthesize six borophenes at higher temperatures,liberating them from the computer virtual world and promoting their development,Borophene has good properties,such as light density,high strength,good flexibility,easy to chemical reactions.Thereby,it is widely used in the synthesis of superconducting materials,battery electrode materials,hydrogen storage materials,gas sensors and hydrogen evolution reaction catalysts.However,borophene with poor stability cannot be directly applied.We need chemical modification methods to improve stability and mechanical properties to ensure follow-up research.The second part introduces the important role of machine learning algorithms in computational chemistry.It can be combined with DFT method and make up for its shortcomings,providing new ideas for the development of theoretical chemistry.High-throughput approaches can also be used to accelerate the design of new materials,guide the synthesis of chemical reactions,and predict and screen materials with specific properties.The second chapter mainly introduces the related theories and applications of first-principles calculations and machine learning methods,as well as the current mainstream software toolkits in computational chemistry and machine learning frameworks.Density functional theory(DFT)is derived from quantum mechanics.Based on the electron density of the system,the ground state particle density is acquired,and the properties of the system are finally obtained.The main idea to solve the problem is to transform the complex multi-particle system problem into a simple single-particle system problem,and use the exchange correlation functional to get the approximate solution of the system.In addition,the current popular machine learning algorithms can be divided into three categories:unsupervised learning,supervised learning and reinforcement learning,which have good applications in various fields.For situations with a small amount of data and simple descriptors,classic machine learning algorithms can often quickly obtain excellent training results.If the amount of data and descriptors are more complex,then deep learning methods can take advantage of them.At the same time,when the machine learning algorithm is trained,the selection of parameters is very important,such as the optimizer,activation function,loss function,number of hidden layers,and number of neurons.Selecting appropriate methods for specific problems in materials research can often get twice the result with half the effort.In the third chapter,the influence of chemical modification on the band structure and stability of borophene was discussed by first-principles calculation.As a kind of two-dimensional materials,borophene has attracted much attention due to its excellent properties.At present,the synthesis of some borophenes has been completed by complex MBE technology under harsh conditions,which greatly promotes its development.In this paper,DFT method was used to study the band gap opening and stability of six experimentally synthesized borophenes.We found that the band gaps of ?3,?,?3,?5 borophenes can be opened by chemical modification methods(hydrogenation,fluorination and chlorination),while ?6 and ?12 borophenes cannot open the band gaps in this way.In addition,the phonon spectrum,AIMD simulation and elastic constant analysis show that the dynamic and thermodynamic stability and mechanical properties of these borophenes have been improved to a certain extent.Especially for ?6 borophene,the band gap is opened and the stability is greatly improved.These results provide guidance and suggestions for the further study of borophene and related semiconductor materials.In Chapter 4,we study the prediction of the dissociation energy of organic carbonyls by using machine learning method.Bond dissociation energy(BDE)is an indicator of the strength of chemical bonds.It has shown great potential in evaluating and screening high-performance materials and catalysts,which is vital in industrial applications.However,it is usually expensive and time-consuming to measure or calculate BDE through conventional experimental or theoretical methods,which greatly prevents the application of BDE to large-scale and high-throughput research.Therefore,there is a great need for a more effective method of estimating BDE.To this end,first-principles calculations and machine learning techniques,including neural networks and random forests,are combined to explore the internal connection between the carbonyl structure and its BDE.We selected the bond length and bond angle information of the carbonyl group as descriptors to train the model,and compared the final prediction results with the DFT calculation results.At the same time,valence electrons are added to convert the data into Coulomb force information between atoms for further discussion.The results show that machine learning can not only effectively reproduce the calculated carbonyl BDE,but also provide guidance for rational design of carbonyl structure to optimize performance.