A Combined First-principles And Machine Learning Investigation On The Electronic Structures Of Two-Dimensional Structures

Since its first isolation in 2004,graphene and related two-dimensional(2D)materials have attracted considered interest in scientific and engineering communities.In the past decade,2D materials have been extensively explored for fields of batteries,catalysis,electronics and photonics among many others.Meanwhile,the unique properties of graphene have also inspired researchers to seek for new members of 2D material family.These new 2D materials have also only proposed to apply in practical electronic devices such as sensors,LEDs and FETs,but also exhibited extraordinary potential for other applications in physics,catalysis chemistry,biomedical and environmental sciences.In recent years,development of first-principles calculations has accelerated design the development of new 2D structures.However,due to the limited space of known compounds,it is necessary to employ other approaches to discover new 2D structures and study their physical properties.In recent years,a rapidly developing branch of artificial intelligence is machine learning,which has been widely employed to solve complex problems involving a large number of combinatorial spaces or nonlinear processes that cannot be solved by traditional methods or can only be solved at high computational cost.In this thesis,the electronic structures of 2D materials is explored by means of machine learning combined with first principles approaches.The main contents of this paper are as follows,(1)By combining first-principles calculations with machine learning,we explored the work function of various 2D materials.The model was established by using the already obtained properties such as atomic number,ion radius,elastic modulus,electron affinity and ionization energy as input characteristics,and by using random forest(RF),extremely randomized trees(ET),gradient boosting regression trees(GBR)and extreme gradient boosting regression(XGBR)as learning methods.The root mean square error(rmse)of the work function predicted by XGBR model in a few seconds was 0.344 e V,which is quite low,and demonstrate that XGBR model is effective to predict the work function of 2D materials.Furthermore,we have also chosen several predicted materials for first-principles calculations,and confirmed the effectiveness of the learning model.(2)We developed two machine learning models using XGBoost classifier to predict the magnetic properties and thermodynamic stability of magnetic 2D materials by using a data set made of 3709 2D materials.A total of 377 features were generated based on element attribute data,and the top 160 and 164 features were selected by the two models according to the order of feature importance.We show that these features were enough to generate the most accurate model without over-fitting.The model was used to predict the magnetic properties and thermodynamic stability of the compounds that do not exist in the training set,which was found to be effective.These results demonstrate that the machine learning is a fast and effective method and can provide useful guidance for the exploration of the electronic properties of various 2D materials as well as material design.