Fig. 2: The accuracy of the energy model (RMSE per atom) and the calculated specific heats for FeCoNiAlTi and MoNbTaW.

a The training error of the EPI model for FeCoNiAlTi. b The validation error of the EPI model for FeCoNiAlTi (0.7/0.3 train-valid splitting). c The training error of the qSRO (Local_SRO) model for FeCoNiAlTi. d The validation error of the qSRO model for FeCoNiAlTi. e The specific heats CV of FeCoNiAlTi were calculated with different supercell sizes. f The specific heats of MoNbTaW were calculated with a one million-atom supercell.