Fig. 3: Summary of density functional theory (DFT) validations of the potential phase transformations from model predictions of generations 1–13. | npj Computational Materials

Fig. 3: Summary of density functional theory (DFT) validations of the potential phase transformations from model predictions of generations 1–13.

From: Discovery of new high-pressure phases – integrating high-throughput DFT simulations, graph neural networks, and active learning

Fig. 3

For generations 5–13, one pair was picked in each 50-GPa transformation pressure interval, but not every bin has predicted transformations. Therefore, the total number of pairs that were picked varies slightly across generations. For generation 4, we did not select any pairs to validate since we noticed that graph neural networks (GNN) can have wrong ground state phase prediction. The pairs with gray diagonal lines are pairs that were not able to converge in the CellRelaxDFT tool because of the 24 hours wall time limitation that is set in the tool. These pairs were run outside of the CellRelaxDFT tool after generation 13 and were not included in any of the training data. Two pairs that we could not converge even with the expensive computational resource used are marked gray on the plot.

Back to article page