Predicting energy and stability of known and hypothetical crystals using graph neural network

ABSTRACT: Summary The discovery of new inorganic materials in unexplored chemical spaces necessitates calculating total energy quickly and with sufficient accuracy. Machine learning models that provide such a capability for both ground-state (GS) and higher-energy structures would be instrumental in accelerated screening. Here, we demonstrate the importance of a balanced training dataset of GS and higher-energy structures to accurately predict total energies using a generic graph neural network architecture. Using Highlights • A GNN is trained to predict total energy of ground-state and high-energy crystals• The importance of a balanced training dataset is demonstrated• The model ranks polymorphic structures of compounds with correct energy ordering The bigger picture Large-scale ab initio calculations combined with advances in structure prediction have been instrumental in inorganic functional materials discovery. Currently, only a small fraction of the vast chemical space of inorganic materials has been discovered. The need for accelerated exploration of uncharted chemical spaces is shared by experimental and computational researchers. However, structure prediction and evaluation of phase stability using ab initio methods is intractable to explore vast search spaces. Here, we demonstrate the importance of a balanced training dataset of ground-state (GS) and higher-energy structures to accurately predict their total energies using a generic graph neural network. We demonstrate that the model satisfactorily ranks the structures in the correct order of their energies for a given composition. Together, these capabilities allow the model to be used for fast prediction of GS structures and phase stability and for the facilitation of new materials discovery. Discovery of new inorganic materials in unexplored chemical spaces necessitates calculating total energy quickly and with sufficient accuracy. Structure prediction and phase stability using ab initio methods is intractable to explore vast search spaces. We demonstrate the importance of a balanced training dataset of ground-state and higher-energy structures to accurately predict their total energies using a generic GNN architecture. We also demonstrate that the model satisfactorily ranks the structures in the correct order of their energies for a given composition.

SUBMITTER: Pandey S 

PROVIDER: S-EPMC8600245 | biostudies-literature |

REPOSITORIES: biostudies-literature