Fig. 5: Flowcharts and results comparison on using ct-UAE trained on different tasks or not for perovskite property prediction. | Nature Communications

Fig. 5: Flowcharts and results comparison on using ct-UAE trained on different tasks or not for perovskite property prediction.

From: Transformer-generated atomic embeddings to enhance prediction accuracy of crystal properties with machine learning

Fig. 5

Ef is the formation energy. MAE is the Mean Absolute Error. R2 is the R-squared value in predicting each property. The prefix None- denotes models that do not use ct-UAE. The prefix CT- indicates models that use ct-UAE. The perfix CTMT@np is models that use ct-UAE trained by n properties. a Schematic representation of the workflow for applying ct-UAEs to predict properties of perovskite materials. When the back-end model is MEGNET10, the MAE for UAE-free case is 0.32 eV/atom. Using the transfer learning strategy with ct-UAE results in an MAE of 0.030 eV/atom, while the MAE for the transfer learning strategy with ct-UAE trained using multi-task learning is 0.021 eV/atom. b, c Predicted formation energy versus target formation energy for the MEGNET10 and CGCNN9 models. The upper part and the right part denotes target and prediction data distribution respectively.

Back to article page