Description
This model is a CHGNet universal potential for the PyTorch Geometric (PyG) backend of MatGL. The weights were directly transferred from the DGL checkpoint materialyze/CHGNet-PES-MatPES-r2SCAN-2025.2.10 — no retraining was performed.
The architecture is a faithful PyG port of the original DGL CHGNet implementation. The DGL implementation has a slight modification from the original PyTorch implementation by adding directed edge updates; this PyG port preserves that modification.
Training dataset
MatPES-r2SCAN-2024.11: Materials Energy Surface dataset that contains off-equilibrium r2SCAN static calculations.
- Train-Val-Test splitting with mp-id: 0.9 - 0.5 - 0.5
- Train set size: 349107
- Validation set size: 19395
- Test set size: 19395
Performance metrics
Training and validation errors
Identical to the source DGL checkpoint (weights are the same):
| partition | Energy (meV/atom) | Force (meV/Å) | stress (GPa) | magmom (μB) |
|---|---|---|---|---|
| Train | 26.36 | 85.7 | 0.359 | 0.067 |
| Validation | 27.51 | 150.5 | 0.705 | 0.066 |
| Test | 30.45 | 156.5 | 0.735 | 0.072 |
PyG vs DGL prediction parity
Evaluated on 10 structures spanning diverse chemistries (MoS, Fe, Mo, Al, NaCl, BaTiO₃, Li₂O, MgO, and their perturbed variants) with both the PyG and DGL models running independently.
| Quantity | Max |ΔPyG − ΔDGL| | |---|---| | Energy/atom (eV) | 0 (exact) | | Forces (eV/Å) | 2.7 × 10⁻⁷ | | Stress (GPa) | 4.2 × 10⁻⁶ | | Magnetic moment (μB) | 0 (exact) |
Non-zero force/stress differences arise from floating-point summation order differences
between DGL and PyG message-passing kernels, not from any model divergence.
All differences are more than two orders of magnitude below the atol = 1e-5 threshold
of the automated parity test.
Usage
import matgl
from matgl.ext.pymatgen import Structure2Graph
import torch
pot = matgl.load_model("BowenD-UCB/CHGNet-PyG-MatPES-r2SCAN-2025.2.10")
pot.eval()
References
Deng, B. et al. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.
Nat. Mach. Intell. 1–11 (2023) doi:10.1038/s42256-023-00716-3.
Date: 2025.2.10
Author: Bowen Deng
Metadata
{
"tags": [
"matgl",
"materials-science",
"graph-neural-network"
],
"license": "BSD-3-Clause",
"author": "Bowen Deng"
}