materialyze/SO3Net-PES-ANI-1x-Subset

Description

This model is a M3GNet potential for 4 elements including H, C, N, O. It has broad applications in the dynamic simulations of organic molecules.

Training dataset

ANI-1x-Subset: 300K MD simulations and Materials Project ground state calculations.

• Training set size: 991735
• Validation set size: 248355
• Test set size: 248355

Performance metrics

Training and validation errors

MAEs of energies, forces and stresses, respectively

• Training: 2.281 meV/atom, 46 meV/A
• Validation: 2.286 meV/atom, 46 meV/A
• Test: 1.596 meV/atom, 35 meV/A

References

Ko, T.W., Deng, B., Nassar, M. et al. Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry. npj Computation Materials 11, 253 (2025).