摘要
Certain perovskite-type oxynitrides have bandgaps suitable for renewable hydrogen production via photocatalytic and photoelectrochemical water splitting under visible light.Understanding the ordering of oxide and nitride anions in these materials is important because this ordering affects their semiconductor properties.However, the numerous possible orderings complicate systematic analyses based on density functional theory(DFT) calculations using defined elemental arrangements.This work shows that anion ordering in large-scale supercells within perovskite-type oxynitrides can be rapidly predicted based on machine learning, using BaNbO2N(capable of oxidizing water under irradiation up to 740 nm) as an example.Machine learning allows the calculation of the total energy of BaNbO2N directly from randomly selected initial atomic placements without costly structural optimization, thus reducing the computational cost by more than 99.99%.Combined with the Metropolis Monte Carlo method, machine learning permits exploration of the stable anion orderings of large supercells without costly DFT calculations.This work therefore demonstrates a means of predicting the properties of functional materials having complex compositions based on the most realistic elemental arrangements in conjunction with reasonable computational loads.
Certain perovskite-type oxynitrides have bandgaps suitable for renewable hydrogen production via photocatalytic and photoelectrochemical water splitting under visible light.Understanding the ordering of oxide and nitride anions in these materials is important because this ordering affects their semiconductor properties.However, the numerous possible orderings complicate systematic analyses based on density functional theory（DFT） calculations using defined elemental arrangements.This work shows that anion ordering in large-scale supercells within perovskite-type oxynitrides can be rapidly predicted based on machine learning, using BaNbO2N（capable of oxidizing water under irradiation up to 740 nm） as an example.Machine learning allows the calculation of the total energy of BaNbO2N directly from randomly selected initial atomic placements without costly structural optimization, thus reducing the computational cost by more than 99.99%.Combined with the Metropolis Monte Carlo method, machine learning permits exploration of the stable anion orderings of large supercells without costly DFT calculations.This work therefore demonstrates a means of predicting the properties of functional materials having complex compositions based on the most realistic elemental arrangements in conjunction with reasonable computational loads.
基金
financially supported by Grants-in-Aid for Scientific Research (A) (no.16H02417)
Young Scientists (A) (no.15H05494) from the Japan Society for the Promotion of Science (JSPS)
partly supported by MEXT as “Priority Issue on Post-K computer” (Development of new fundamental technologies for highefficiency energy creation, conversion/storage and use)
