Fig. 5: Performance of CGCNN for the prediction of DFT formation energies with and without transfer learning.
From: Shotgun crystal structure prediction using machine-learned formation energies
The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) for test instances are shown on each parity plot. a Results of global model for the prediction of relaxed formation energies of 90 benchmark crystals (orange). b Histogram of DFT formation energies of relaxed and randomly generated pre-relaxed structures. c, d Prediction of pre-relaxed formation energies of 100 randomly generated conformations for each of the 90 benchmark systems (c) without and (d) with fine-tuning of the pretrained global energy prediction model.
