README.md

---
license: cc-by-4.0
---

## Model Summary

This model card provides a DPA3 model[1] trained on the OMol25[2] dataset. We provide one model:

- `DPA3-Omol-Large` – 12 layers, large-scale model for broad molecular chemistry

The model is trained with **charge** and **spin** as input frame parameters, following the OMol25 dataset convention. Here **spin** refers to the **spin multiplicity** (2S+1), not the spin quantum number S. Users can specify `charge` and `spin` when running simulations; if not specified, defaults of `charge=0` and `spin=1` (singlet) are used.

The model is compatible with **DeePMD-kit v3.1.3**. For other installation options, please visit the [Releases page](https://github.com/deepmodeling/deepmd-kit/releases/tag/v3.1.3) to download the off-line package for v3.1.3, and refer to the [official documentation](https://docs.deepmodeling.com/projects/deepmd/en/v3.1.3/install/easy-install.html) for off-line installation instructions.

## Usage

### Model Evaluation

Evaluate the model through the dp test command line:

```bash
dp --pt test -m DPA3-Omol-Large.pt -s path_to_your_system
```

### ASE Calculator

You can directly use the following Python code for prediction or optimization with standard ASE calculator.

**Charge and spin** can be explicitly specified via `fparam` keyword in `atoms.info`. Note that `spin` here means **spin multiplicity** (2S+1). If not set, the default values `charge=0` and `spin=1` (singlet) will be used.

```python
## Compute potential energy
from ase import Atoms
from deepmd.calculator import DP as DPCalculator

dp = DPCalculator("DPA3-Omol-Large.pt")

# Example: ethanol molecule
ethanol = Atoms(
    "C2H6O",
    positions=[
        (-0.7472, -0.0575, 0.0000),
        ( 0.7209,  0.0178, 0.0000),
        ( 1.1431,  1.4297, 0.0000),
        (-1.1576, -1.0720, 0.0000),
        (-1.1267,  0.4548, -0.8932),
        (-1.1267,  0.4548,  0.8932),
        ( 1.0797, -0.5050, -0.8946),
        ( 1.0797, -0.5050,  0.8946),
        ( 2.1108,  1.4520,  0.0000),
    ],
    cell=[100, 100, 100],
)

# Specify charge and spin multiplicity (optional)
# If not set, defaults are charge=0, spin=1 (singlet)
ethanol.info.update(
  {"fparam": [0.0, 1.0]} # charge=0, spin multiplicity=1 (singlet)
)

ethanol.calc = dp
print(ethanol.get_potential_energy())
print(ethanol.get_forces())

## Run BFGS structure optimization
from ase.optimize import BFGS

dyn = BFGS(ethanol)
dyn.run(fmax=1e-6)
print(ethanol.get_positions())
```

### LAMMPS

Use LAMMPS for molecular dynamics calculation with the DPA3 model, you first need to freeze the *.pt model into a *.pth model using the following command:

```bash
dp --pt freeze -c DPA3-Omol-Large.pt -o DPA3-Omol-Large.pth
```

Then you can make the following modifications in the LAMMPS script to call the DeePMD-kit interface (also see `potential.md`).

**Charge and spin** are provided via the `fparam` keyword in the order of **charge, spin** (spin = spin multiplicity, 2S+1). If `fparam` is not specified, the default values `0.0 1.0` (charge=0, spin multiplicity=1, i.e. singlet) will be used.

```bash
# With explicit charge and spin multiplicity (e.g., charge=2, multiplicity=1)
pair_style      deepmd DPA3-Omol-Large.pth fparam 2.0 1.0
pair_coeff      * * C H O
```

```bash
# Without fparam: defaults to charge=0, spin multiplicity=1
pair_style      deepmd DPA3-Omol-Large.pth
pair_coeff      * * C H O
```

For more details on the `fparam` keyword, see the [DeePMD-kit LAMMPS documentation](https://docs.deepmodeling.com/projects/deepmd/en/stable/third-party/lammps-command.html#pair-style-deepmd).

## Training Dataset

The model is trained on the **Open Molecules 2025 (OMol25)** dataset[2], a large-scale resource for molecular chemistry ML models introduced by Meta FAIR. OMol25 comprises over **100 million** DFT single-point calculations at the **ωB97M-V/def2-TZVPD** level of theory.

Key characteristics of OMol25:

- **83 elements** across the periodic table
- **~83M unique molecular systems**, including small molecules, biomolecules, metal complexes, and electrolytes
- System sizes up to **350 atoms** (50 on average)
- Diverse charge states (−10 to +10) and spin multiplicities (1 to 11)
- Explicit solvation, conformers, and reactive structures

The dataset is organized into four major domains:

- **Biomolecules:** protein–ligand, protein–protein, and nucleic acid interactions extracted from BioLiP2 and other structural databases, sampled via classical MD
- **Metal Complexes:** diverse monometallic transition metal, main group metal, and lanthanide systems with varied ligands and spin states, generated using the Architector package
- **Electrolytes:** aqueous and non-aqueous solutions, ionic liquids, and molten salts, sampled via MD (including Ring Polymer MD for nuclear quantum effects) and electrolyte reactivity networks
- **Community:** recomputed existing datasets (ANI-2X, Transition-1X, SPICE2, GEOM, etc.) at consistent ωB97M-V/def2-TZVPD level of theory, plus interpolated reactivity datasets

Compositional splitting ensures that validation and test sets contain out-of-distribution molecular formulas relative to training data.

## Training Details

We train the **DPA3** model in its large (12-layer) configuration, truncated within LiGS order 2.

### Model configuration


| Parameter | Value |
| --------- | ----- |
| `n_dim`   | 256   |
| `e_dim`   | 256   |
| `a_dim`   | 256   |
| `nlayers` | 12    |

### Training setup

- **Engine:** DeePMD-kit (`v3.1.0` required)
- **Batch size:** `auto:2048` (DeePMD-kit automatic batchsize)
- **Hardware:** 32 × NVIDIA A800 GPUs
- **Training steps:** 2 million steps
- **Learning rate schedule:** Cosine annealing
- **Cutoff radii and neighbor selections:**
  - `e_rcut = 6.0`, `e_rcut_smth = 5.3`, `e_sel = 30`
  - `a_rcut = 4.5`, `a_rcut_smth = 4.0`, `a_sel = 15`

Other hyperparameters and training details can be found in the DPA3 paper[1].

## Performance

### Accuracy on OMol25 Validation Set

We report energy and force errors on the OMol25 validation set. All values are in **meV** (energy per atom) and **meV/Å** (force).


| Model               | Energy MAE/atom (meV) | Energy RMSE/atom (meV) | Force MAE (meV/Å) | Force RMSE (meV/Å) |
| ------------------- | :-------------------: | :--------------------: | :----------------: | :-----------------: |
| MACE-OMol-L         |         1.917         |         11.727         |       10.690       |       63.754       |
| **DPA3-Omol-Large** |         1.328         |         11.347         |       12.362       |       62.934       |

### Accuracy on LAMBench

We evaluate DPA3-Omol-Large on molecule property calculation tasks from **LAMBench**[3]. The following results compare **DPA3-Omol-Large (ours)** with DPA-3.2-5M and MACE-OMol-L.

#### Ligand Binding


| Model               | RMSE (kcal/mol) | MAE (kcal/mol) |
| ------------------- | :-------------: | :------------: |
| DPA-3.2-5M          |      3.40      |      6.90      |
| MACE-OMol-L         |      1.75      |      0.84      |
| **DPA3-Omol-Large** |      1.29      |      0.65      |

#### TorsionNet500


| Model               | MAEB (kcal/mol) | MAE (kcal/mol) | RMSE (kcal/mol) | NAHB_h |
| ------------------- | :-------------: | :------------: | :-------------: | :----: |
| DPA-3.2-5M          |      0.47      |      0.29      |      0.43      |   50   |
| MACE-OMol-L         |      0.23      |      0.14      |      0.23      |   9   |
| **DPA3-Omol-Large** |      0.24      |      0.16      |      0.25      |   13   |

#### Wiggle150


| Model               | RMSE (kcal/mol) | MAE (kcal/mol) |
| ------------------- | :-------------: | :------------: |
| DPA-3.2-5M          |      1.54      |      1.19      |
| MACE-OMol-L         |      1.18      |      0.89      |
| **DPA3-Omol-Large** |      1.25      |      0.94      |

#### Reaction Barrier


| Model               | RMSE (kcal/mol) | MAE (kcal/mol) |
| ------------------- | :-------------: | :------------: |
| DPA-3.2-5M          |      12.37      |      6.30      |
| MACE-OMol-L         |      3.53      |      2.12      |
| **DPA3-Omol-Large** |      12.42      |      3.36      |

## Reference

[1] Duo Zhang, Anyang Peng, Chun Cai, Wentao Li, Yuanchang Zhou, Jinzhe Zeng, Mingyu Guo et al. "A Graph Neural Network for the Era of Large Atomistic Models." *arXiv preprint arXiv:2506.01686* (2025).

[2] Daniel S. Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G. Taylor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia et al. "The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models." *arXiv preprint arXiv:2505.08762* (2025).

[3] Anyang Peng, Chun Cai, Mingyu Guo, Duo Zhang, Chengqian Zhang, Wanrun Jiang, Yinan Wang et al. "LAMBench: a benchmark for large atomistic models." *npj Computational Materials* (2026).