Fig. 5: DFT validation on DefiNet’s accuracy, efficiency, and scalability.
From: Modeling crystal defects using defect informed neural networks
a Comparison of the number of DFT ionic steps required to relax structures starting from the initial unrelaxed configurations and from the DefiNet-predicted structures for low-density defects. The steady residual ionic steps against the defect complexity are indicated by a horizontal black solid line. The sample ID is sorted based on the number of ionic steps required by the unrelaxed structures for better observation. b Residual ionic steps for five randomly selected defect structures from the 50 samples across different supercell sizes, starting from DefiNet-predicted configurations. Only a single reference run is shown for unrelaxed structures due to the high computational cost of initiating DFT relaxation from unrelaxed configurations. The steady residual ionic steps against the structural size are indicated by a horizontal black solid line. c Comparison of DFT CPU core hours on large supercells using unrelaxed and DefiNet-predicted configurations. Due to the extremely high computational cost associated with the unrelaxed structure of the 16 × 16 supercell size with 770 atoms, only one sample was selected as an example for this experiment.
