Research On Finding Unusual Structures With Machine Learning

Machine learning（ML）has found wide applications on fields like vision recognition and machine translation.In result years,it has also been used in materials science to build models to predict physical and chemical properties of materials from their structures.Such scheme has been demonstrated in previous literatures focusing on properties such as band gap,ionic conductivity,melting temperature,etc.However,there are only a few works that pay attention to the building of ML models,and purely statistical methods are used in these works,which requires heavy calculation.Another problem is that while ML can be used to find common relationships between structures and properties,there are always exceptions that will shed some new insights about the underlying physics.Thus,a study of these exceptional cases would be quite useful.In this work,data from MaterialGo（MG）database,a new database constructed by our school,are used to build a ML model to predict Heyd-Scuseria-Ernzerhof（HSE）band gaps from crystal structures.By iteratively analyzing predicted results,finding common features among poorly predicted structures and using these to improve the model,we build a model with relatively lighter calculations that shows decent performance.Then,we analyze structures that still have a large prediction error and find some unusual structural units or other abnormities among them.Afterwards,we choose one of them,AgO₂F,for investigation,where both Ag³⁺and O₂^2-exist.It is found that the states near the band edges in AgO₂F are dominated by the non-bonding O-2p orbitals,and the weak hybridization between these non-bonding orbitals and the orbitals of Ag eventually results in an unexpected decrease in the band gap.ML model in this work is built with open-sourced packages scikit-learn and Lightgbm.other data processing,especially feature building,is done with python scripts written by ourselves.PWmat is used for theoretical calculation.Vesta is used for crystal structure visualization and analysis.