Pretrained models#

So far, we have released multiple pretrained SevenNet models. Each model has various hyperparameters and training sets, leading to different levels of accuracy and speed. Please read the descriptions below carefully and choose the model that best suits your purpose. We provide the F1 score, and RMSD for the WBM dataset, along with $\kappa_{\mathrm{SRME}}$ from phononDB and CPS (Combined Performance Score). For details on these metrics and performance comparisons with other pretrained models, please visit Matbench Discovery.

These models can be used as interatomic potentials in LAMMPS and loaded through the ASE calculator using each model’s keywords. Please refer to the ASE calculator section for instructions on loading a model via the ASE calculator. Additionally, keywords can be used in other parts of SevenNet, such as sevenn inference, sevenn get_model, and the checkpoint section in input.yaml for fine-tuning.

Acknowledgments: The models trained on MPtrj were supported by the Neural Processing Research Center program at Samsung Advanced Institute of Technology, part of Samsung Electronics Co., Ltd. The computations for training models were carried out using the Samsung SSC-21 cluster.

Note

Multiple inference tasks are available for multi-fidelity architecture models, SevenNet-Omni and SevenNet-MF-ompa. Each task is designed to produce results that are consistent with the DFT settings used in the corresponding training datasets. For example, mpa is trained on the combined MPtrj and sAlex datasets and is used for evaluating Matbench Discovery, while omat24 is trained on the OMat24 dataset. For detailed information on the DFT settings, please refer to the original publications of each dataset.

Note

We supports cuEquivariance and FlashTP kernels for tensor-product acceleration. Pass these options to SevenNetCalculator using enable_cueq=True or enable_flash=True when available.

SevenNet pretrained models overview#
Model name	Trained dataset(s)	Model architecture	Single/multi-task	Supporting tasks
SevenNet-Omni (Recommended)	15 open ${\mathit{ab}}$ ${\mathit{initio}}$ datasets	$l_{\mathrm{max}}=3$ $N_{\mathrm{layer}}=5$	Multi-task	`mpa` (PBE+U) `matpes_r2scan` (r²SCAN) `omol25_low` (ωB97M-V) and 10 more tasks
SevenNet-MF-ompa	MPtrj, sAlex, OMat24	$l_{\mathrm{max}}=3$ $N_{\mathrm{layer}}=5$	Multi-task	`mpa` `omat24`
SevenNet-omat	OMat24	$l_{\mathrm{max}}=3$ $N_{\mathrm{layer}}=5$	Single-task	N/A
SevenNet-l3i5	MPtrj	$l_{\mathrm{max}}=3$ $N_{\mathrm{layer}}=5$	Single-task	N/A
SevenNet-0	MPtrj	$l_{\mathrm{max}}=2$ $N_{\mathrm{layer}}=5$	Single-task	N/A

SevenNet-Omni#

Model keywords: 7net-Omni | SevenNet-Omni

This is our recommended pretrained model

This model exploits cross-domain knowledge transfer strategies within a multi-task training framework to train simultaneously on the 15 open ab initio datasets, covering a wide material space including crystals, molecules, and surfaces.

It is currently our best pretrained model, achieving state-of-the-art accuracy across diverse material domains at the PBE level, while also providing high-fidelity tasks such as r²SCAN and ωB97M-V. For detailed information on the training datasets, knowledge-transfer strategies and comprehensive benchmark results, please refer to the paper.

Representative tasks for each fidelity#
Fidelity	Task name
PBE(+U)	`mpa`
r²SCAN	`matpes_r2scan`
ωB97M-V	`omol25_low`

Also consider omat24 or matpes_pbe tasks for more accurate PBE-level descriptions of high-force configurations. Note that matpes_pbe is trained on PBE level of theory, without incorporating the Hubbard U correction. All available tasks and corresponding fidelities are listed below. Tasks may differ even at the same fidelity level due to variations in computational protocols across datasets.

List of available tasks and corresponding fidelities#
Task name	Fidelity
`mpa`	PBE(+U)
`omat24`	PBE(+U)
`matpes_pbe`	PBE
`oc20`	RPBE
`oc22`	PBE(+U)
`odac23`	PBE-D3
`omol25_low`	ωB97M-V
`omol25_high`	ωB97M-V*
`spice`	ωB97M
`qcml`	PBE0
`pet_mad`	PBEsol
`mp r2scan`	r²SCAN
`matpes_r2scan`	r²SCAN

* For only high-spin configurations

from sevenn.calculator import SevenNetCalculator
calc = SevenNetCalculator(
    model="/path/to/7net-omni",
    modal='mpa',
    enable_cueq=False,
    enable_flash=False
)

When using the command-line interface of SevenNet, include the task as --modal ${task} option to select the desired task. Run sevenn cp 7net-omni to see an overview of this model.

Matbench Discovery#

CPS	F1	$\kappa_{\mathrm{SRME}}$	RMSD
0.849	0.889	0.265	0.0639

SevenNet-MF-ompa#

Model keywords: 7net-mf-ompa | SevenNet-mf-ompa

This model leverages multi-fidelity learning to train simultaneously on the MPtrj, sAlex, and OMat24 datasets. This model achieves a high ranking on the Matbench Discovery leaderboard. Our evaluations show that it outperforms other models on most tasks, except for the isolated molecule energy task, where it performs slightly worse than SevenNet-l3i5.

from sevenn.calculator import SevenNetCalculator
# "mpa" refers to the MPtrj + sAlex task, used for evaluating Matbench Discovery.
calc = SevenNetCalculator('7net-mf-ompa', modal='mpa')  # Use modal='omat24' for OMat24-trained task weights.

When using the command-line interface of SevenNet, include the --modal mpa or --modal omat24 option to select the desired task.

Matbench Discovery#

CPS	F1	$\kappa_{\mathrm{SRME}}$	RMSD
0.845	0.901	0.317	0.064

Detailed instructions for multi-fidelity learning

Download link for fully detailed checkpoint

SevenNet-omat#

Model keywords: 7net-omat | SevenNet-omat

This model was trained exclusively on the OMat24 dataset. It achieves high performance in $\kappa_{\mathrm{SRME}}$ on Matbench Discovery, but its F1 score is unavailable due to a difference in the POTCAR version. Like SevenNet-MF-ompa, this model outperforms SevenNet-l3i5 on most tasks, except for the isolated molecule energy.

Download link for fully detailed checkpoint.

Matbench Discovery#

$\kappa_{\mathrm{SRME}}$: 0.221

SevenNet-l3i5#

Model keywords: 7net-l3i5 | SevenNet-l3i5

This model increases the maximum spherical harmonic degree ($l_{\mathrm{max}}$) to 3, compared to SevenNet-0, which has an $l_{\mathrm{max}}$ of 2. While l3i5 offers improved accuracy for various systems, it is approximately four times slower than SevenNet-0.

Matbench Discovery#

CPS	F1	$\kappa_{\mathrm{SRME}}$	RMSD
0.714	0.760	0.550	0.085

SevenNet-0#

Model keywords:: 7net-0 | SevenNet-0 | 7net-0_11Jul2024 | SevenNet-0_11Jul2024

This model is one of our earliest pretrained models. Although we recommend using newer and more accurate models, it can still be useful in certain cases due to its shortest inference time. The model was trained on the MPtrj and is loaded as the default pretrained model in the ASE calculator.

Matbench Discovery#

F1	$\kappa_{\mathrm{SRME}}$
0.67	0.767

Model name	Trained dataset(s)	Model architecture	Single/multi-task	Supporting tasks
SevenNet-Omni (Recommended)	15 open \({\mathit{ab}}\) \({\mathit{initio}}\) datasets	\(l_{\mathrm{max}}=3\) \(N_{\mathrm{layer}}=5\)	Multi-task	`mpa` (PBE+U) `matpes_r2scan` (r²SCAN) `omol25_low` (ωB97M-V) and 10 more tasks
SevenNet-MF-ompa	MPtrj, sAlex, OMat24	\(l_{\mathrm{max}}=3\) \(N_{\mathrm{layer}}=5\)	Multi-task	`mpa` `omat24`
SevenNet-omat	OMat24	\(l_{\mathrm{max}}=3\) \(N_{\mathrm{layer}}=5\)	Single-task	N/A
SevenNet-l3i5	MPtrj	\(l_{\mathrm{max}}=3\) \(N_{\mathrm{layer}}=5\)	Single-task	N/A
SevenNet-0	MPtrj	\(l_{\mathrm{max}}=2\) \(N_{\mathrm{layer}}=5\)	Single-task	N/A

Pretrained models#

SevenNet-Omni#

Matbench Discovery#

SevenNet-MF-ompa#

Matbench Discovery#

SevenNet-omat#

Matbench Discovery#

SevenNet-l3i5#

Matbench Discovery#

SevenNet-0#

Matbench Discovery#

This Page