MODELS.md

Pretrained OCP models

This page summarizes all the pretrained models released as part of the Open Catalyst Project. All models were trained using this codebase.

• Open Catalyst 2020 (OC20)
  • S2EF models optimized for EFwT
  • S2EF models optimized for force
  • IS2RE models
• Open Catalyst 2022 (OC22)
  • S2EF total models
• Open Direct Air Capture 2023 (ODAC23)
  • S2EF models
  • IS2RE Direct models

---

Open Catalyst 2020 (OC20)

• All configurations for these models are available in the configs/ directory.
• All of these models are trained on various splits of the OC20 S2EF / IS2RE datasets. For details, see https://arxiv.org/abs/2010.09990 and https://github.com/Open-Catalyst-Project/ocp/blob/main/DATASET.md.
• All OC20 models are trained on adsorption energies, i.e. the DFT total energies minus the clean surface and gas phase adsorbate energies. For details on how to train models on OC20 total energies, please read the referencing section here.

S2EF models: optimized for EFwT

Model	Split	Download	val ID force MAE (eV / Å)	val ID EFwT
CGCNN	200k	checkpoint | config	0.08	0%
CGCNN	2M	checkpoint | config	0.0673	0.01%
CGCNN	20M	checkpoint | config	0.065	0%
CGCNN	All	checkpoint | config	0.0684	0.01%
DimeNet	200k	checkpoint	0.0693	0.01%
DimeNet	2M	checkpoint	0.0576	0.02%
SchNet	200k	checkpoint | config	0.0743	0%
SchNet	2M	checkpoint | config	0.0737	0%
SchNet	20M	checkpoint | config	0.0568	0.03%
SchNet	All	checkpoint | config	0.0494	0.12%
DimeNet++	200k	checkpoint | config	0.0741	0%
DimeNet++	2M	checkpoint | config	0.0595	0.01%
DimeNet++	20M	checkpoint | config	0.0511	0.06%
DimeNet++	All	checkpoint | config	0.0444	0.12%
SpinConv	2M	checkpoint | config	0.0329	0.18%
SpinConv	All	checkpoint | config	0.0267	1.02%
GemNet-dT	2M	checkpoint | config	0.0257	1.10%
GemNet-dT	All	checkpoint | config	0.0211	2.21%
PaiNN	All	checkpoint | config | scale file	0.0294	0.91%
GemNet-OC	2M	checkpoint | config | scale file	0.0225	2.12%
GemNet-OC	All	checkpoint | config | scale file	0.0179	4.56%
GemNet-OC	All+MD	checkpoint | config | scale file	0.0173	4.72%
GemNet-OC-Large	All+MD	checkpoint | config | scale file	0.0164	5.34%
SCN	2M	checkpoint | config	0.0216	1.68%
SCN-t4-b2	2M	checkpoint | config	0.0193	2.68%
SCN	All+MD	checkpoint | config	0.0160	5.08%
eSCN-L4-M2-Lay12	2M	checkpoint | config	0.0191	2.55%
eSCN-L6-M2-Lay12	2M	checkpoint | config	0.0186	2.66%
eSCN-L6-M2-Lay12	All+MD	checkpoint | config	0.0161	4.28%
eSCN-L6-M3-Lay20	All+MD	checkpoint | config	0.0139	6.64%
EquiformerV2 (83M)	2M	checkpoint | config	0.0167	4.26%
EquiformerV2 (31M)	All+MD	checkpoint | config	0.0142	6.20%
EquiformerV2 (153M)	All+MD	checkpoint | config	0.0126	8.90%

S2EF models: optimized for force only

Model	Split	Download	val ID force MAE
SchNet	All	checkpoint	0.0443
DimeNet++	All	checkpoint | config	0.0334
DimeNet++-Large	All	checkpoint | config	0.02825
DimeNet++	20M+Rattled	checkpoint	0.0614
DimeNet++	20M+MD	checkpoint	0.0594

IS2RE models

Model	Split	Download	val ID energy MAE
CGCNN	10k	checkpoint | config	0.9881
CGCNN	100k	checkpoint | config	0.682
CGCNN	All	checkpoint | config	0.6199
DimeNet	10k	checkpoint	1.0117
DimeNet	100k	checkpoint	0.6658
DimeNet	All	checkpoint	0.5999
SchNet	10k	checkpoint | config	1.059
SchNet	100k	checkpoint | config	0.7137
SchNet	All	checkpoint | config	0.6458
DimeNet++	10k	checkpoint | config	0.8837
DimeNet++	100k	checkpoint | config	0.6388
DimeNet++	All	checkpoint | config	0.5639
PaiNN	All	checkpoint | config | scale file	0.5728

The Open Catalyst 2020 (OC20) dataset is licensed under a Creative Commons Attribution 4.0 License.

Please consider citing the following paper in any research manuscript using the OC20 dataset or pretrained models, as well as the original paper for each model:

@article{ocp_dataset,
    author = {Chanussot*, Lowik and Das*, Abhishek and Goyal*, Siddharth and Lavril*, Thibaut and Shuaibi*, Muhammed and Riviere, Morgane and Tran, Kevin and Heras-Domingo, Javier and Ho, Caleb and Hu, Weihua and Palizhati, Aini and Sriram, Anuroop and Wood, Brandon and Yoon, Junwoong and Parikh, Devi and Zitnick, C. Lawrence and Ulissi, Zachary},
    title = {Open Catalyst 2020 (OC20) Dataset and Community Challenges},
    journal = {ACS Catalysis},
    year = {2021},
    doi = {10.1021/acscatal.0c04525},
}

Open Catalyst 2022 (OC22)

• All configurations for these models are available in the configs/oc22 directory.
• All of these models are trained on various splits of the OC22 S2EF / IS2RE datasets. For details, see https://arxiv.org/abs/2206.08917 and https://github.com/Open-Catalyst-Project/ocp/blob/main/DATASET.md.
• All OC22 models released here are trained on DFT total energies, in contrast to the OC20 models listed above, which are trained on adsorption energies.

S2EF-Total models

Model	Training	Download	val ID force MAE	val ID energy MAE
GemNet-dT	OC22	checkpoint | config	0.032	1.127
GemNet-OC	OC22	checkpoint | config	0.030	0.563
GemNet-OC	OC20+OC22	checkpoint | config	0.027	0.483
GemNet-OC (trained with enforce_max_neighbors_strictly=False, #467)	OC20+OC22	checkpoint | config	0.027	0.458
GemNet-OC	OC20->OC22	checkpoint | config	0.030	0.417
EquiformerV2 ($\lambda_E$=4, $\lambda_F$=100)	OC22	checkpoint | config	0.023	0.447

The Open Catalyst 2022 (OC22) dataset is licensed under a Creative Commons Attribution 4.0 License.

Please consider citing the following paper in any research manuscript using the OC22 dataset or pretrained models, as well as the original paper for each model:

@article{oc22_dataset,
    author = {Tran*, Richard and Lan*, Janice and Shuaibi*, Muhammed and Wood*, Brandon and Goyal*, Siddharth and Das, Abhishek and Heras-Domingo, Javier and Kolluru, Adeesh and Rizvi, Ammar and Shoghi, Nima and Sriram, Anuroop and Ulissi, Zachary and Zitnick, C. Lawrence},
    title = {The Open Catalyst 2022 (OC22) dataset and challenges for oxide electrocatalysts},
    journal = {ACS Catalysis},
    year={2023},
}

Open Direct Air Capture 2023 (ODAC23)

• All config files for the ODAC23 models are available in the configs/odac directory.

S2EF models

Model	Checkpoint	Config
SchNet	checkpoint	config
DimeNet++	checkpoint	config
PaiNN	checkpoint	config
GemNet-OC	checkpoint	config
eSCN	checkpoint	config
EquiformerV2	checkpoint	config
EquiformerV2 (Large)	checkpoint	config

IS2RE Direct models

Model	Checkpoint	Config
Gemnet-OC (Direct)	checkpoint	config
eSCN (Direct)	checkpoint	config
EquiformerV2 (Direct)	checkpoint	config

The models in the table above were trained to predict relaxed energy directly. Relaxed energies can also be predicted by running structural relaxations using the S2EF models from the previous section.

IS2RS

The IS2RS is solved by running structural relaxations using the S2EF models from the prior section.

The Open DAC 2023 (ODAC23) dataset is licensed under a Creative Commons Attribution 4.0 License.

Please consider citing the following paper in any research manuscript using the ODAC23 dataset:

@article{odac23_dataset,
    author = {Anuroop Sriram and Sihoon Choi and Xiaohan Yu and Logan M. Brabson and Abhishek Das and Zachary Ulissi and Matt Uyttendaele and Andrew J. Medford and David S. Sholl},
    title = {The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture},
    year = {2023},
    journal={arXiv preprint arXiv:2311.00341},
}