G-NequIP: Data-Efficient Generalization of Equivariant Neural Network Potentials across Diverse Organic Chemical Space

This dataset accompanies the manuscript: "G-NequIP: A Data-Efficient Equivariant Neural Network Potential for Generalized Molecular Simulations" All training data, benchmark geometries and energies, pre-trained model weights (. pth files), and analysis scripts used in the manuscript are included. The deposit is organized into eight sections: - Dataₚool — ~3. 3 million molecular structures used for training, drawn from GDB-11, ChEMBL, QM9, Hutchison, MNSol, S66×8, and amino acid databases- NNPₜraining — Training datasets (by source and cumulative) in extended XYZ format with cohesive energies and atomic forces; pre-trained G-NequIP model weights (. pth files) - PESₛcans — Potential energy surface torsion scan data comparing G-NequIP, AIMNet2, ANI-2x, and DFT reference energies- NMSbenchmark — Normal mode sampling benchmark geometries and single-point energies for molecules with ≤10, 10–20, and 20–30 heavy atoms- conformergeneration — Conformer ensembles generated with DeepConf, Auto3D, and loqi for benchmark molecules- CCSᵥalidation — Collision cross-section benchmark data for 20 compounds including MobCal-MPI inputs and outputs- MDVDOS — Molecular dynamics trajectories and vibrational density of states data- solvationfreeₑnergies — Implicit solvation (SMD) free energy results with G-NequIP-smdW Software repositories: DeepConf (https: //github. com/otayfuroglu/DeepConf) and G-NequIP training code (https: //github. com/otayfuroglu/deepPotential).