Data for "Aligning Heterogeneous DFT Datasets: A Graph Neural Network Approach to Cross-Functional Formation Energies"

This dataset accompanies a study of cross-functional formation-energy prediction and correction between PBE and r²SCAN. It contains the processed paired formation-energy dataset, GPTFF-develop prediction records, sanitized GPTFF and CHGNet training and evaluation results, machine-readable source data for Figs. 1–5, and voltage-calculation records for LiCoO₂ and LiMn₂O₄. During data curation, compounds containing Tb, Sm, or Dy were excluded because of their unusually large deviations between the calculated and experimental formation energies. LiVRh₂ was additionally excluded because of the anomalously large difference between the two density-functional approximations. The voltage-calculation archive contains generated Li/vacancy configurations, retained PBE relaxation results, PBE and r²SCAN static-calculation records, model-corrected energies, convergence information, excluded-candidate records, and the data used to construct the voltage profiles. Nonconverged relaxations and failed static calculations are explicitly documented. The figure-source archive provides the machine-readable inputs required by the public figure-generation scripts. SHA-256 checksums are supplied in MANIFEST.sha256 to verify file integrity. VASP executables and licensed PAW datasets are not redistributed. The corresponding code, trained correction-model checkpoints, environment specification, and reproduction instructions are available at:https://github.com/Jack1620/cross-functional-formation-energy